Apple Patent | Point Cloud Compression With Multi-Layer Projection

Patent: Point Cloud Compression With Multi-Layer Projection

Publication Number: 20190156519

Publication Date: 20190523

Applicants: Apple

Abstract

A system comprises an encoder configured to compress attribute information and/or spatial for a point cloud and/or a decoder configured to decompress compressed attribute and/or spatial information for the point cloud. To compress the attribute and/or spatial information, the encoder is configured to convert a point cloud into an image based representation. Also, the decoder is configured to generate a decompressed point cloud based on an image based representation of a point cloud. The encoder is configured project the point cloud on to patch planes to compress the point cloud, and supports multiple layered patch planes. For example, some point clouds may have a depth, and points at different depths may be assigned to different layered patch planes.

[0001] This application claims benefit of priority to U.S. Provisional Application Ser. No. 62/590,191, entitled “Point Cloud Compression with Multi-Lay Projection,” filed Nov. 22, 2017, and which is incorporated herein by reference in its entirety.

BACKGROUND

Technical Field

[0002] This disclosure relates generally to compression and decompression of point clouds comprising a plurality of points, each having associated spatial information and attribute information.

Description of the Related Art

[0003] Various types of sensors, such as light detection and ranging (LIDAR) systems, 3-D-cameras, 3-D scanners, etc. may capture data indicating positions of points in three dimensional space, for example positions in the X, Y, and Z planes. Also, such systems may further capture attribute information in addition to spatial information for the respective points, such as color information (e.g. RGB values), texture information, intensity attributes, reflectivity attributes, motion related attributes, modality attributes, or various other attributes. In some circumstances, additional attributes may be assigned to the respective points, such as a time-stamp when the point was captured. Points captured by such sensors may make up a “point cloud” comprising a set of points each having associated spatial information and one or more associated attributes. In some circumstances, a point cloud may include thousands of points, hundreds of thousands of points, millions of points, or even more points. Also, in some circumstances, point clouds may be generated, for example in software, as opposed to being captured by one or more sensors. In either case, such point clouds may include large amounts of data and may be costly and time-consuming to store and transmit.

SUMMARY OF EMBODIMENTS

[0004] In some embodiments, a system includes one or more sensors configured to capture points that collectively make up a point cloud, wherein each of the points comprises spatial information identifying a spatial location of the respective point and attribute information defining one or more attributes associated with the respective point.

[0005] The system also includes an encoder configured to compress the attribute and/or spatial information of the points. To compress the attribute and/or spatial information, the encoder is configured to determine, for the point cloud, a plurality of patches each corresponding to portions of the point cloud, wherein each patch comprises points with surface normal vectors that deviate from one another less than a threshold amount. The encoder is further configured to, for each patch, generate another patch image comprising the set of points corresponding to the patch projected onto a patch plane and generate a patch image comprising depth information for the set of points corresponding to the patch, wherein the depth information represents depths of the points in a direction perpendicular to the patch plane.

[0006] For example, the patch image corresponding to the patch projected onto a patch plane may depict the points of the point cloud included in the patch in two directions, such as an X and Y direction. The points of the point cloud may be projected onto a patch plane approximately perpendicular to a normal vector, normal to a surface of the point cloud at the location of the patch. Also, for example, the patch image comprising depth information for the set of points included in the patch may depict depth information, such as depth distances in a Z direction. To depict the depth information, the depth patch image may include a parameter that varies in intensity based on the depth of points in the point cloud at a particular location in the patch image. For example, the patch image depicting depth information may have a same shape as the patch image representing points projected onto the patch plane. However, the depth information patch image may be an image comprising image attributes, such as one or more colors, that vary in intensity, wherein the intensity of the one or more image attributes corresponds to a depth of the point cloud at a location in the patch image where the image attribute is displayed in the patch image depicting depth. For example, points that are closer to the patch plane may be encoded as darker values in the patch image depicting depth and points that are further away from the patch plane may be encoded as brighter values in the patch image depicting depth, for example in a monochromatic patch image depicting depth. Thus, the depth information patch image when aligned with other patch images representing points projected onto the patch plane may indicate the relative depths of the points projected onto the patch plane, based on respective image attribute intensities at locations in the depth patch image that correspond to locations of the points in the other patch images comprising point cloud points projected onto the patch plane.

[0007] In some embodiments, points of a point cloud may be in a same or nearly same location when projected onto a patch plane. For example, the point cloud might have a depth such that some points are in the same location relative to the patch plane, but at different depths. In such embodiments, multiple patches may be generated for different layers of the point cloud. In some embodiments, subsequent layered patches may encode differences between a previous layer, such that the subsequent layers do not repeat the full amount of data encoded in the previous layer(s). Thus, subsequent layers may have significantly smaller sizes than initial layers.

[0008] The encoder is further configured to pack generated patch images (including a depth patch image and, optionally, one or more additional patch images for one or more other attributes) for each of the determined patches into one or more image frames and encode the one or more image frames. In some embodiments, the encoder may utilize various image or video encoding techniques to encode the one or more image frames. For example, the encoder may utilize a video encoder in accordance with the High Efficiency Video Coding (HEVC/H.265) standard or other suitable standards such as, the Advanced Video Coding (AVC/H.265) standard, the AOMedia Video 1 (AV1) video coding format produced by the Alliance for Open Media (AOM), etc. In some embodiments, the encoder may utilize an image encoder in accordance with a Motion Picture Experts Group (MPEG), a Joint Photography Experts Group (JPEG) standard, an International Telecommunication Union-Telecommunication standard (e.g. ITU-T standard), etc.

[0009] In some embodiments, colors of patch images packed into image frames may be converted into a different color space or may be sub-sampled to further compress the image frames. For example, in some embodiments a 4:4:4 R’G’B’ color space may be converted into a 4:2:0 YCbCr color space. Additionally, a color conversion process may determine an optimal luma value and corresponding chroma values. For example, a an optimal luma value may be selected that reduces a converted size of the fame image while minimizing distortion of the decompressed point cloud colors as compared to the original non-compressed point cloud. In some embodiments, an iterative approach may be used to determine an optimal luma value. In other embodiments, one or more optimization equations may be applied to determine an optimal luma and corresponding chroma values.

[0010] Such a system may further account for distortion caused by projecting the point cloud onto patches and packing the patches into image frames. Additionally, such a system may account for distortion caused by video encoding and decoding the image frames comprising packed patches. To do this, a closed-loop color conversion module may take as an input a reference point cloud original color and a video compressed image frame comprising packed patches, wherein the packed patches of the image frame have been converted from a first color space to a second color space. The closed-loop color conversion module may decompress the compressed image frame using a video decoder and furthermore reconstruct the point cloud using the decompressed image frames. The closed-loop color conversion module may then determine color values for points of the decompressed point cloud based on attribute and/or texture information included in the decompressed patches of the decompressed image frames (in the converted color space). The closed-loop color conversion module may then compare the down sampled and up sampled colors of the reconstructed point cloud to the colors of the original non-compressed point cloud. Based on this comparison, the closed-loop color conversion module may then adjust one or more parameters used to convert the image frames from the original color space to the second color space, wherein the one or more parameters are adjusted to improve quality of the final decompressed point cloud colors and to reduce a size of the compressed point cloud.

[0011] In some embodiments, a decoder is configured to receive one or more encoded image frames comprising patch images for a plurality of patches of a compressed point cloud, wherein, for each patch, the one or more encoded image frames comprise: a patch image comprising a set of points of the patch projected onto a patch plane and a patch image comprising depth information for the set of points of the patch, wherein the depth information indicates depths of the points of the patch in a direction perpendicular to the patch plane. In some embodiments, a depth patch image may be packed into an image frame with other attribute patch images. For example, a decoder may receive one or more image frames comprising packed patch images as generated by the encoder described above.

[0012] The decoder is further configured to decode the one or more encoded image frames comprising the patch images. In some embodiments, the decoder may utilize a video decoder in accordance with the High Efficiency Video Coding (HEVC) standard or other suitable standards such as, the Advanced Video Coding (AVC) standard, the AOMedia Video 1 (AV1) video coding format, etc. In some embodiments, the decoder may utilize an image decoder in accordance with a Motion Picture Experts Group (MPEG) or a Joint Photography Experts Group (JPEG) standard, etc.

[0013] The decoder is further configured to determine, for each patch, spatial information for the set of points of the patch based, at least in part, on the patch image comprising the set of points of the patch projected onto the patch plane and the patch image comprising the depth information for the set of points of the patch, and generate a decompressed version of the compressed point cloud based, at least in part, on the determined spatial information for the plurality of patches and the attribute information included in the patches.

[0014] In some embodiments, a method includes receiving one or more encoded image frames comprising patch images for a plurality of patches of a compressed point cloud, wherein, for each patch, the one or more encoded image frames comprise: a patch image comprising a set of points of the patch projected onto a patch plane and a patch image comprising depth information for the set of points of the patch, wherein the depth information indicates depths of the points of the patch in a direction perpendicular to the patch plane. The method further includes decoding the one or more encoded image frames comprising the patch images. In some embodiments, decoding may be performed in accordance with the High Efficiency Video Coding (HEVC) standard or other suitable standards such as, the Advanced Video Coding (AVC) standard, an AOMedia Video 1 (AV1) video coding format, etc. In some embodiments, decoding may be performed in accordance with a Motion Picture Experts Group (MPEG) or a Joint Photography Experts Group (JPEG) standard, etc.

[0015] The method further includes determining, for each patch, spatial information for the set of points of the patch based, at least in part, on the patch image comprising the set of points of the patch projected onto the patch plane and the patch image comprising the depth information for the set of points of the patch and generating a decompressed version of the compressed point cloud based, at least in part, on the determined spatial information for the plurality of patches.

[0016] In some embodiments, a non-transitory computer-readable medium stores program instructions that, when executed by one or more processors, cause the one or more processors to implement an encoder as described herein to compress attribute information of a point cloud.

[0017] In some embodiments, a non-transitory computer-readable medium stores program instructions that, when executed by one or more processors, cause the one or more processors to implement a decoder as described herein to decompress attribute information of a point cloud.

BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIG. 1 illustrates a system comprising a sensor that captures information for points of a point cloud and an encoder that compresses spatial information and attribute information of the point cloud, where the compressed spatial and attribute information is sent to a decoder, according to some embodiments.

[0019] FIG. 2A illustrates components of an encoder for encoding intra point cloud frames, according to some embodiments.

[0020] FIG. 2B illustrates components of a decoder for decoding intra point cloud frames, according to some embodiments.

[0021] FIG. 2C illustrates components of an encoder for encoding inter point cloud frames, according to some embodiments.

[0022] FIG. 2D illustrates components of a decoder for decoding inter point cloud frames, according to some embodiments.

[0023] FIG. 2E illustrates components of a closed-loop color conversion module, according to some embodiments.

[0024] FIG. 2F illustrates an example process for determining a quality metric for a point cloud upon which an operation has been performed, according to some embodiments.

[0025] FIG. 3A illustrates an example patch segmentation process, according to some embodiments.

[0026] FIG. 3B illustrates an example image frame comprising packed patch images and padded portions, according to some embodiments.

[0027] FIG. 3C illustrates an example image frame comprising patch portions and padded portions, according to some embodiments.

[0028] FIG. 3D illustrates a point cloud being projected onto multiple projections, according to some embodiments.

[0029] FIG. 3E illustrates a point cloud being projected onto multiple parallel projections, according to some embodiments.

[0030] FIG. 4A illustrates a process for compressing attribute and spatial information of a point cloud, according to some embodiments.

[0031] FIG. 4B illustrates a process for decompressing attribute and spatial information of a point cloud, according to some embodiments.

[0032] FIG. 4C illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a point cloud, according to some embodiments.

[0033] FIG. 4D illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a moving or changing point cloud, according to some embodiments.

[0034] FIG. 4E illustrates a decoder receiving image frames comprising patch images, patch information, and an occupancy map, and generating a decompressed representation of a point cloud, according to some embodiments.

[0035] FIG. 4F illustrates an encoder, adjusting encoding based on one or more masks for a point cloud, according to some embodiments.

[0036] FIG. 4G illustrates a decoder, adjusting decoding based on one or more masks for a point cloud, according to some embodiments.

[0037] FIG. 4H illustrates more detail regarding compression of an occupancy map, according to some embodiments.

[0038] FIG. 4I illustrates example blocks and traversal patterns for compressing an occupancy map, according to some embodiments.

[0039] FIG. 5 illustrates compressed point cloud information being used in a 3-D telepresence application, according to some embodiments.

[0040] FIG. 6 illustrates compressed point cloud information being used in a virtual reality application, according to some embodiments.

[0041] FIG. 7 illustrates an example computer system that may implement an encoder or decoder, according to some embodiments.

[0042] This specification includes references to “one embodiment” or “an embodiment.” The appearances of the phrases “in one embodiment” or “in an embodiment” do not necessarily refer to the same embodiment. Particular features, structures, or characteristics may be combined in any suitable manner consistent with this disclosure.

[0043] “Comprising.” This term is open-ended. As used in the appended claims, this term does not foreclose additional structure or steps. Consider a claim that recites: “An apparatus comprising one or more processor units … . ” Such a claim does not foreclose the apparatus from including additional components (e.g., a network interface unit, graphics circuitry, etc.).

[0044] “Configured To.” Various units, circuits, or other components may be described or claimed as “configured to” perform a task or tasks. In such contexts, “configured to” is used to connote structure by indicating that the units/circuits/components include structure (e.g., circuitry) that performs those task or tasks during operation. As such, the unit/circuit/component can be said to be configured to perform the task even when the specified unit/circuit/component is not currently operational (e.g., is not on). The units/circuits/components used with the “configured to” language include hardware–for example, circuits, memory storing program instructions executable to implement the operation, etc. Reciting that a unit/circuit/component is “configured to” perform one or more tasks is expressly intended not to invoke 35 U.S.C. .sctn. 112(f), for that unit/circuit/component. Additionally, “configured to” can include generic structure (e.g., generic circuitry) that is manipulated by software and/or firmware (e.g., an FPGA or a general-purpose processor executing software) to operate in manner that is capable of performing the task(s) at issue. “Configure to” may also include adapting a manufacturing process (e.g., a semiconductor fabrication facility) to fabricate devices (e.g., integrated circuits) that are adapted to implement or perform one or more tasks.

[0045] “First,” “Second,” etc. As used herein, these terms are used as labels for nouns that they precede, and do not imply any type of ordering (e.g., spatial, temporal, logical, etc.). For example, a buffer circuit may be described herein as performing write operations for “first” and “second” values. The terms “first” and “second” do not necessarily imply that the first value must be written before the second value.

[0046] “Based On.” As used herein, this term is used to describe one or more factors that affect a determination. This term does not foreclose additional factors that may affect a determination. That is, a determination may be solely based on those factors or based, at least in part, on those factors. Consider the phrase “determine A based on B.” While in this case, B is a factor that affects the determination of A, such a phrase does not foreclose the determination of A from also being based on C. In other instances, A may be determined based solely on B.

DETAILED DESCRIPTION

[0047] As data acquisition and display technologies have become more advanced, the ability to capture point clouds comprising thousands or millions of points in 2-D or 3-D space, such as via LIDAR systems, has increased. Also, the development of advanced display technologies, such as virtual reality or augmented reality systems, has increased potential uses for point clouds. However, point cloud files are often very large and may be costly and time-consuming to store and transmit. For example, communication of point clouds over private or public networks, such as the Internet, may require considerable amounts of time and/or network resources, such that some uses of point cloud data, such as real-time uses, may be limited. Also, storage requirements of point cloud files may consume a significant amount of storage capacity of devices storing the point cloud files, which may also limit potential applications for using point cloud data.

[0048] In some embodiments, an encoder may be used to generate a compressed point cloud to reduce costs and time associated with storing and transmitting large point cloud files. In some embodiments, a system may include an encoder that compresses attribute and/or spatial information of a point cloud file such that the point cloud file may be stored and transmitted more quickly than non-compressed point clouds and in a manner that the point cloud file may occupy less storage space than non-compressed point clouds. In some embodiments, compression of attributes of points in a point cloud may enable a point cloud to be communicated over a network in real-time or in near real-time. For example, a system may include a sensor that captures attribute information about points in an environment where the sensor is located, wherein the captured points and corresponding attributes make up a point cloud. The system may also include an encoder that compresses the captured point cloud attribute information. The compressed attribute information of the point cloud may be sent over a network in real-time or near real-time to a decoder that decompresses the compressed attribute information of the point cloud. The decompressed point cloud may be further processed, for example to make a control decision based on the surrounding environment at the location of the sensor. The control decision may then be communicated back to a device at or near the location of the sensor, wherein the device receiving the control decision implements the control decision in real-time or near real-time. In some embodiments, the decoder may be associated with an augmented reality system and the decompressed attribute information may be displayed or otherwise used by the augmented reality system. In some embodiments, compressed attribute information for a point cloud may be sent with compressed spatial information for points of the point cloud. In other embodiments, spatial information and attribute information may be separately encoded and/or separately transmitted to a decoder.

[0049] In some embodiments, a system may include a decoder that receives one or more sets of point cloud data comprising compressed attribute information via a network from a remote server or other storage device that stores the one or more point cloud files. For example, a 3-D display, a holographic display, or a head-mounted display may be manipulated in real-time or near real-time to show different portions of a virtual world represented by point clouds. In order to update the 3-D display, the holographic display, or the head-mounted display, a system associated with the decoder may request point cloud data from the remote server based on user manipulations of the displays, and the point cloud data may be transmitted from the remote server to the decoder and decoded by the decoder in real-time or near real-time. The displays may then be updated with updated point cloud data responsive to the user manipulations, such as updated point attributes.

[0050] In some embodiments, a system, may include one or more LIDAR systems, 3-D cameras, 3-D scanners, etc., and such sensor devices may capture spatial information, such as X, Y, and Z coordinates for points in a view of the sensor devices. In some embodiments, the spatial information may be relative to a local coordinate system or may be relative to a global coordinate system (for example, a Cartesian coordinate system may have a fixed reference point, such as a fixed point on the earth, or may have a non-fixed local reference point, such as a sensor location).

[0051] In some embodiments, such sensors may also capture attribute information for one or more points, such as color attributes, reflectivity attributes, velocity attributes, acceleration attributes, time attributes, modalities, and/or various other attributes. In some embodiments, other sensors, in addition to LIDAR systems, 3-D cameras, 3-D scanners, etc., may capture attribute information to be included in a point cloud. For example, in some embodiments, a gyroscope or accelerometer, may capture motion information to be included in a point cloud as an attribute associated with one or more points of the point cloud. For example, a vehicle equipped with a LIDAR system, a 3-D camera, or a 3-D scanner may include the vehicle’s direction and speed in a point cloud captured by the LIDAR system, the 3-D camera, or the 3-D scanner. For example, when points in a view of the vehicle are captured they may be included in a point cloud, wherein the point cloud includes the captured points and associated motion information corresponding to a state of the vehicle when the points were captured.

Example System Arrangement

[0052] FIG. 1 illustrates a system comprising a sensor that captures information for points of a point cloud and an encoder that compresses attribute information of the point cloud, where the compressed attribute information is sent to a decoder, according to some embodiments.

[0053] System 100 includes sensor 102 and encoder 104. Sensor 102 captures a point cloud 110 comprising points representing structure 106 in view 108 of sensor 102. For example, in some embodiments, structure 106 may be a mountain range, a building, a sign, an environment surrounding a street, or any other type of structure. In some embodiments, a captured point cloud, such as captured point cloud 110, may include spatial and attribute information for the points included in the point cloud. For example, point A of captured point cloud 110 comprises X, Y, Z coordinates and attributes 1, 2, and 3. In some embodiments, attributes of a point may include attributes such as R, G, B color values, a velocity at the point, an acceleration at the point, a reflectance of the structure at the point, a time stamp indicating when the point was captured, a string-value indicating a modality when the point was captured, for example “walking”, or other attributes. The captured point cloud 110 may be provided to encoder 104, wherein encoder 104 generates a compressed version of the point cloud (compressed attribute information 112) that is transmitted via network 114 to decoder 116. In some embodiments, a compressed version of the point cloud, such as compressed attribute information 112, may be included in a common compressed point cloud that also includes compressed spatial information for the points of the point cloud or, in some embodiments, compressed spatial information and compressed attribute information may be communicated as separate sets of data.

[0054] In some embodiments, encoder 104 may be integrated with sensor 102. For example, encoder 104 may be implemented in hardware or software included in a sensor device, such as sensor 102. In other embodiments, encoder 104 may be implemented on a separate computing device that is proximate to sensor 102.

Example Intra-Frame Encoder

[0055] FIG. 2A illustrates components of an encoder for encoding intra point cloud frames, according to some embodiments. In some embodiments, the encoder described above in regard to FIG. 1 may operate in a similar manner as encoder 200 described in FIGS. 2A and encoder 250 described in FIG. 2C.

[0056] The encoder 200 receives uncompressed point cloud 202 and generates compressed point cloud information 204. In some embodiments, an encoder, such as encoder 200, may receive the uncompressed point cloud 202 from a sensor, such as sensor 102 illustrated in FIG. 1, or, in some embodiments, may receive the uncompressed point cloud 202 from another source, such as a graphics generation component that generates the uncompressed point cloud in software, as an example.

[0057] In some embodiments, an encoder, such as encoder 200, includes decomposition into patches module 206, packing module 208, spatial image generation module 210, texture image generation module 212, and attribute information generation module 214. In some embodiments, an encoder, such as encoder 200, also includes image frame padding module 216, closed loop color conversion module 217, video compression module 218 and multiplexer 224. In addition, in some embodiments an encoder, such as encoder 200, may include an occupancy map compression module, such as occupancy map compression module 220, and an auxiliary patch information compression module, such as auxiliary patch information compression module 222. In some embodiments, an encoder, such as encoder 200, converts a 3D point cloud into an image-based representation along with some meta data (e.g., occupancy map and patch info) necessary to convert the compressed point cloud back into a decompressed point cloud.

[0058] In some embodiments, the conversion process decomposes the point cloud into a set of patches (e.g., a patch is defined as a contiguous subset of the surface described by the point cloud), which may be overlapping or not, such that each patch may be described by a depth field with respect to a plane in 2D space. More details about the patch decomposition process are provided below with regard to FIGS. 3A-3C.

[0059] After or in conjunction with the patches being determined for the point cloud being compressed, a 2D sampling process is performed in planes associated with the patches. The 2D sampling process may be applied in order to approximate each patch with a uniformly sampled point cloud, which may be stored as a set of 2D patch images describing the geometry/texture/attributes of the point cloud at the patch location. The “Packing” module 208 may store the 2D patch images associated with the patches in a single (or multiple) 2D images, referred to herein as “image frames.” In some embodiments, a packing module, such as packing module 208, may pack the 2D patch images such that the packed 2D patch images do not overlap (even though an outer bounding box for one patch image may overlap an outer bounding box for another patch image). Also, the packing module may pack the 2D patch images in a way that minimizes non-used images pixels of the image frame.

[0060] In some embodiments, “Geometry/Texture/Attribute generation” modules, such as modules 210, 212, and 214, generate 2D patch images associated with the geometry/texture/attributes, respectively, of the point cloud at a given patch location. As noted before, a packing process, such as performed by packing module 208, may leave some empty spaces between 2D patch images packed in an image frame. Also, a padding module, such as image frame padding module 216, may fill in such areas in order to generate an image frame that may be suited for 2D video and image codecs.

[0061] In some embodiments, an occupancy map (e.g., binary information describing for each pixel or block of pixels whether the pixel or block of pixels are padded or not) may be generated and compressed, for example by occupancy map compression module 220. The occupancy map may be sent to a decoder to enable the decoder to distinguish between padded and non-padded pixels of an image frame.

[0062] Note that other metadata associated with patches may also be sent to a decoder for use in the decompression process. For example, patch information indicating sizes and shapes of patches determined for the point cloud and packed in an image frame may be generated and/or encoded by an auxiliary patch-information compression module, such as auxiliary patch-information compression module 222. In some embodiments one or more image frames may be encoded by a video encoder, such as video compression module 218. In some embodiments, a video encoder, such as video compression module 218, may operate in accordance with the High Efficiency Video Coding (HEVC) standard or other suitable video encoding standard. In some embodiments, encoded video images, encoded occupancy map information, and encoded auxiliary patch information may be multiplexed by a multiplexer, such as multiplexer 224, and provided to a recipient as compressed point cloud information, such as compressed point cloud information 204.

[0063] In some embodiments, an occupancy map may be encoded and decoded by a video compression module, such as video compression module 218. This may be done at an encoder, such as encoder 200, such that the encoder has an accurate representation of what the occupancy map will look like when decoded by a decoder. Also, variations in image frames due to lossy compression and decompression may be accounted for by an occupancy map compression module, such as occupancy map compression module 220, when determining an occupancy map for an image frame. In some embodiments, various techniques may be used to further compress an occupancy map, such as described in FIGS. 4H-4I.

Example Intra-Frame Decoder

[0064] FIG. 2B illustrates components of a decoder for decoding intra point cloud frames, according to some embodiments. Decoder 230 receives compressed point cloud information 204, which may be the same compressed point cloud information 204 generated by encoder 200. Decoder 230 generates reconstructed point cloud 246 based on receiving the compressed point cloud information 204. In some embodiments, a decoder, such as decoder 230, includes a de-multiplexer 232, a video decompression module 234, an occupancy map decompression module 236, and an auxiliary patch-information decompression module 238. Additionally a decoder, such as decoder 230 includes a point cloud generation module 240, which reconstructs a point cloud based on patch images included in one or more image frames included in the received compressed point cloud information, such as compressed point cloud information 204. In some embodiments, a decoder, such as decoder 203, further comprises a smoothing filter, such as smoothing filter 244. In some embodiments, a smoothing filter may smooth incongruences at edges of patches, wherein data included in patch images for the patches has been used by the point cloud generation module to recreate a point cloud from the patch images for the patches. In some embodiments, a smoothing filter may be applied to the pixels located on the patch boundaries to alleviate the distortions that may be caused by the compression/decompression process.

Example Inter-Frame Encoder

[0065] FIG. 2C illustrates components of an encoder for encoding inter point cloud frames, according to some embodiments. An inter point cloud encoder, such as inter point cloud encoder 250, may encode an image frame, while considering one or more previously encoded/decoded image frames as references.

[0066] In some embodiments, an encoder for inter point cloud frames, such as encoder 250, includes a point cloud re-sampling module 252, a 3-D motion compensation and delta vector prediction module 254, a spatial image generation module 256, a texture image generation module 258, and an attribute image generation module 260. In some embodiments, an encoder for inter point cloud frames, such as encoder 250, may also include an image padding module 262, a closed-loop color conversion module 263, and a video compression module 264. An encoder for inter point cloud frames, such as encoder 250, may generate compressed point cloud information, such as compressed point cloud information 266. In some embodiments, the compressed point cloud information may reference point cloud information previously encoded by the encoder, such as information from or derived from one or more reference image frames. In this way an encoder for inter point cloud frames, such as encoder 250, may generate more compact compressed point cloud information by not repeating information included in a reference image frame, and instead communicating differences between the reference frames and a current state of the point cloud.

[0067] In some embodiments, an encoder, such as encoder 250, may be combined with or share modules with an intra point cloud frame encoder, such as encoder 200. In some embodiments, a point cloud re-sampling module, such as point cloud re-sampling module 252, may resample points in an input point cloud image frame in order to determine a one-to-one mapping between points in patches of the current image frame and points in patches of a reference image frame for the point cloud. In some embodiments, a 3D motion compensation & delta vector prediction module, such as a 3D motion compensation & delta vector prediction module 254, may apply a temporal prediction to the geometry/texture/attributes of the resampled points of the patches. The prediction residuals may be stored into images, which may be padded and compressed by using video/image codecs. In regard to spatial changes for points of the patches between the reference frame and a current frame, a 3D motion compensation & delta vector prediction module 254, may determine respective vectors for each of the points indicating how the points moved from the reference frame to the current frame. A 3D motion compensation & delta vector prediction module 254, may then encode the motion vectors using different image parameters. For example, changes in the X direction for a point may be represented by an amount of red included at the point in a patch image that includes the point. In a similar manner, changes in the Y direction for a point may be represented by an amount of blue included at the point in a patch image that includes the point. Also, in a similar manner, changes in the Z direction for a point may be represented by an amount of green included at the point in a patch image that includes the point. In some embodiments, other characteristics of an image included in a patch image may be adjusted to indicate motion of points included in the patch between a reference frame for the patch and a current frame for the patch.

Example Inter-Frame Decoder

[0068] FIG. 2D illustrates components of a decoder for decoding inter point cloud frames, according to some embodiments. In some embodiments, a decoder, such as decoder 280, includes a video decompression module 270, an inverse 3D motion compensation and inverse delta prediction module 272, a point cloud generation module 274, and a smoothing filter 276. In some embodiments, a decoder, such as decoder 280 may be combined with a decoder, such as decoder 230, or may share some components with the decoder, such as a video decompression module and/or smoothing filter. In decoder 280, the video/image streams are first decoded, then an inverse motion compensation and delta prediction procedure may be applied. The obtained images are then used in order to reconstruct a point cloud, which may be smoothed as described previously to generate a reconstructed point cloud 282.

Closed-Loop Color Conversion

[0069] In some embodiments, an encoder and/or decoder for a point cloud may further include a color conversion module to convert color attributes of a point cloud from a first color space to a second color space. In some embodiments, color attribute information for a point cloud may be more efficiently compressed when converted to a second color space.

[0070] FIG. 2E illustrates components of a closed-loop color conversion module, according to some embodiments. The closed-loop color conversion module 290 illustrated in FIG. 2E may be a similar closed-loop color conversion module as closed-loop color conversion modules 217 and 263 illustrated in FIGS. 2A and 2C.

[0071] In some embodiments, a closed-loop color conversion module, such as closed-loop color conversion module 290 receives a compressed point cloud from a video encoder, such as video encoder 218 illustrated in FIG. 2A or video encoder 264 illustrated in FIG. 2C. Additionally, a closed-loop color conversion module, such as closed-loop color conversion module 290, may receive attribute information about an original non-compressed point cloud, such as color values of points of the point cloud prior to being down-sampled, up-sampled, color converted, etc. Thus, a closed-loop color conversion module may receive a compressed version of a point cloud such as a decoder would receive and also a reference version of the point cloud before any distortion was introduced into the point cloud due to sampling, compression, or color conversion.

[0072] In some embodiments, a closed-loop color conversion module, such as closed-loop color conversion module 290, may include a video compression module, such as video decompression module 280, and a geometry reconstruction module, such as geometry reconstruction module 282. A video decompression module may decompress one or more video encoded image frames to result in decompressed image frames each comprising one or more patch images packed into the image frames. A geometry reconstruction module, such as geometry reconstruction module 282 may then generate a reconstructed point cloud geometry. A re-coloring module, such as re-coloring module 286 may then determine colors for points in the point cloud based on the determined reconstructed geometry. For example, in some embodiments, a nearest neighbor approach or other approach may be used to determine estimated color values for points of the point cloud based on sub-sampled color information, wherein a color value is not explicitly encoded for each point of the point cloud. Because there may be losses during the patching process, compression process, decompression process, and geometry reconstruction process, the geometry of the points in the reconstructed point cloud may not be identical to the geometry in the original point cloud. Due to this discrepancy color compression techniques that rely on geometrical relationships between points to encode color values may result in colors that are slightly different when decoded and decompressed than the original colors. For example, if a color is to be determine based on color values of the nearest neighboring points, a change in geometry may cause a different nearest neighbor to be selected to determine the color value for the point at the decoder than was selected to encode a residual value at the encoder. Thus distortion may be added to the decoded decompressed point cloud.

[0073] If a color space conversion module does not account for this distortion that takes place when converting a point cloud into patches packed in an image frame and that takes place when encoding the image frames, the color space conversion module may select less than optimum color conversion parameters, such as luma and chroma values. For example, optimum color conversion parameters that cause a packed image frame in a first color space to closely match the packed image frame converted into a second color space may be different than optimum color conversion parameters when upstream and downstream distortions are accounted for.

[0074] In order to account for such distortions, a texture/attribute image color space conversion and re-sampling module, such as module 288, may take into account a difference between the “re-created” color values from re-coloring module 288 and the original color values from the original non-compressed reference point cloud when determining color conversion parameters for converting an image frame from a first color space, such as R’G’B’ 4:4:4 to YCbCr 4:2:0, for example. Thus, the color-converted and re-sampled texture/attribute images provided to video encoder 218 and 264, as shown in FIG. 2E may take into account distortion introduced at any stage of compression and decompression of a point cloud, and may utilize optimum color conversion parameters taking into account such distortion.

[0075] Such methods may result in considerably reduced distortion when reconstructing the point cloud representation, while maintaining the high compressibility characteristics of the 4:2:0 signal.

[0076] In some embodiments, conversion from 4:4:4 R’G’B’ to a 4:2:0 YCbCr representation is performed using a 3.times.3 matrix conversion of the form:

[ Y ’ Cb Cr ] = [ a YR a YG a YB a CbR a CbG a CbB a CrR a CrG a CrB ] [ R ’ G ’ B ’ ] ##EQU00001##

[0077] In the above matrix, Y’ is the luma component and Cb and Cr are the chroma components. The values of R’, G’, and B’ correspond to the red, green, and blue components respectively, after the application of a transfer function that is used to exploit the psycho-visual characteristics of the signal. The coefficients aYR through acrB are selected according to the relationship of the red, green, and blue components to the CIE 1931 XYZ color space. Furthermore, the Cb and Cr components are also related to Y’ in the following manner:

Cb = B ’ - Y ’ alpha with alpha = 2 * ( 1 - a YB ) ##EQU00002## Cr = R ’ - Y ’ beta with beta = 2 * ( 1 - a YR ) ##EQU00002.2##

with also the following relationships:

a CbR = - a YR 2 * ( 1 - a YB ) ##EQU00003## a CbR = - a YG 2 * ( 1 - a YB ) ##EQU00003.2## a CbB = 0.5 ##EQU00003.3## a CrR = 0.5 ##EQU00003.4## a CrR = - a YG 2 * ( 1 - a YR ) ##EQU00003.5## a CrB = - a YB 2 * ( 1 - a YR ) ##EQU00003.6##

[0078] The process described above is followed by a 2.times. down-sampling horizontally and vertically of the chroma components, resulting in chroma components that are 4 times smaller, in terms of overall number of samples, 2.times. smaller horizontally and 2.times. smaller vertically, compared to those of luma. Such a process can help not only with compression but also with bandwidth and processing complexity of the YCbCr 4:2:0 signals.

[0079] In using such an approach quantization for the color components, as well as the down sampling and up sampling processes for the chroma components, may introduce distortion that could impact the quality of the reconstructed signals especially in the R’G’B’ but also in the XYZ (CIE 1931 domains). However, a closed loop conversion process, where the chroma and luma values are generated while taking into account such distortions, may considerably improve quality.

[0080] In a luma adjustment process, for example, the chroma components may be converted using the above formulations, additionally a down sampling and up sampling may be performed given certain reference filtering mechanisms. Afterwards, using the reconstructed chroma samples, an appropriate luma value may be computed that would result in minimal distortion for the luminance Y component in the CIE 1931 XYZ space. Such luma value may be derived through a search process instead of a direct computation method as provided above. Refinements and simplifications of this method may include interpolative techniques to derive the luma value.

[0081] Projected point cloud images can also benefit from similar strategies for 4:2:0 conversion. For example, closed loop color conversion, including luma adjustment methods may be utilized in this context. That is, instead of converting point cloud data by directly using the 3.times.3 matrix above and averaging all neighboring chroma values to generate the 4:2:0 chroma representation for the projected image, one may first project point cloud data/patches using the R’G’B’ representation on a 4:4:4 grid. For this new image one may then convert to the YCbCr 4:2:0 representation while using a closed loop optimization such as the luma adjustment method. Assuming that the transfer characteristics function is known, e.g. BT.709, ST 2084 (PQ), or some other transfer function as well as the color primaries of the signal, e.g. BT.709 or BT.2020, an estimate of the luminance component Y may be computed before the final conversion. Then the Cb and Cr components may be computed, down sampled and up sampled using more sophisticated filters. This may then be followed with the computation of the Y’ value that would result in a luminance value Yrecon that would be as close as possible to Y. If distortion in the RGB domain is of higher distortion, a Y’ value that minimizes the distortion for R’, G’, and B’ jointly, could be considered instead.

[0082] For point cloud data, since geometry may also be altered due to lossy compression, texture distortion may also be impacted. In particular, overall texture distortion may be computed by first determining for each point in the original and reconstructed point clouds their closest point in the reconstructed and original point clouds respectively. Then the RGB distortion may be computed for those matched points and accumulated across the entire point cloud image. This means that if the geometry was altered due to lossy compression, the texture distortion would also be impacted. Given that the texture may have been distorted, it may be desirable to consider geometry during closed loop conversion of chroma.

[0083] In some embodiments, the geometry is modified so that the relative sampling density in a given region of the point cloud is adjusted to be similar to other regions of the point cloud. Here the relative sampling density is defined as density of original points relative to the uniform 2D sampling grid.

[0084] Because the relative sampling density can vary within a given patch, this information can be used to guide the patch decomposition process as described above in regard to occupancy maps and auxiliary information, where patch approximation is used to determine local geometry. Furthermore, this information can be used to guide encoding parameters to achieve more uniform quality after compression. If a local region has higher relative sampling density, the encoder may code that region better through a variety of means. The variety of means may include: variable block size decision, Quantization Parameters (QPs), quantization rounding, de-blocking, shape adaptive offset (SAO) filtering, etc.

[0085] In some embodiments, the geometry information is first compressed according to a target bitrate or quality, and then it is reconstructed before generating the texture projected image. Then, given the reconstructed geometry, the closest point in the reconstructed point cloud is determined that corresponds to each point in the original point cloud. The process may be repeated for all points in the reconstructed point cloud by determining their matched points in the original point cloud. It is possible that some points in the reconstructed point cloud may match multiple points in the original point cloud, which would have implications in the distortion computation. This information may be used in the closed loop/luma adjustment method so as to ensure a minimized texture distortion for the entire point cloud. That is, the distortion impact to the entire point cloud of a sample Pr at position (x,y,z) in the reconstructed point cloud can be computed (assuming the use of MSE on YCbCr data for the computation):

D(Pr)=Doriginal(Pr)+Dreconstructed(Pr)

D(Pr)=Sum_matching(((Y_pr-Y_or(i)) 2+(Cb_pr-Cb_or(i)) 2)+sqrt((Y_pr-Y_or) 2+(Cb_pr-Cb_or) 2+(Cr_pr-Cr_or) 2)

[0086] In the above equation, Y_pr, Cb_pr, and Cr_pr are the luma and chroma information of point Pr, Y_or(i), Cb_or(i), and Cr_or(i) correspond to the luma and chroma information of all the points that were found to match the geometry location of point Pr from the original image, and Y_or, Cb_or, and Cr_or is the point that matches the location of point Pr in the original as seen from the reconstructed image.

[0087] If the distortion computation in the context of closed loop conversion/luma adjustment utilizes D(Pr), then better performance may be achieved since it not only optimizes projected distortion, but also point cloud distortion. Such distortion may not only consider luma and chroma values, but may instead or additionally consider other color domain components such as R, G, or B, luminance Y, CIE 1931 x and y, CIE 1976 u’ and v’, YCoCg, and the ICtCp color space amongst others.

[0088] If geometry is recompressed a different optimal distortion point may be possible. In that case, it might be appropriate to redo the conversion process once again.

[0089] In some embodiments, texture distortion, as measured as described below, can be minimized as follows: [0090] Let (Q(j)).sub.i.di-elect cons.{1, … , N} and (P.sub.rec(i)).sub.i.di-elect cons.{1, … , N.sub.rec.sub.} be the original and the reconstructed geometries, respectively. [0091] Let N and N.sub.rec be the number of points in the original and the reconstructed point clouds, respectively. [0092] For each point P.sub.rec(i) in the reconstructed point cloud, let Q*(i) be its nearest neighbor in the original point cloud and R(Q*(i), G (Q*(i), and B (Q*(i) the RGB values associated with Q*(i). [0093] For each point P.sub.rec(i) in the reconstructed point cloud, let .sup.+(i)=(Q.sup.+(i, h)).sub.h.di-elect cons.{1, … , H(i)} be the set of point in the original point cloud that share P.sub.rec(i) as their nearest neighbor in the reconstructed point cloud. Note that .sup.+(i) could be empty or could have one or multiple elements. [0094] If .sup.+(i) is empty, then the RGB values R(Q*(i)), G(Q*(i)), and B(Q*(i)) are associated with the point P.sub.rec. [0095] If .sup.+(i) is not empty, then proceed as follows: [0096] Virtual RGB values, denoted R(.sup.+(i)), G(.sup.+(i)), and B(.sup.+(i)), are computed as follows:

[0096] R ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) R ( Q + ( i , h ) ) ##EQU00004## G ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) G ( Q + ( i , h ) ) ##EQU00004.2## B ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) B ( Q + ( i , h ) ) ##EQU00004.3## [0097] Note that R(.sup.+(i)), G(.sup.+(i)), and B(.sup.+(i)) correspond to the average RGB values of the points of .sup.+(i). [0098] The final RGB values R(P.sub.rec(i)), G(P.sub.rec(i)), and B(P.sub.rec(i)) are obtained by applying the following linear interpolation: [0099] R(P.sub.rec(i))=w R(.sup.+(i))+(1-w) R(Q*(i)) [0100] G(P.sub.rec(i))=w R(.sup.+(i))+(1-w) G(Q*(i)) [0101] B(P.sub.rec(i))=w R(.sup.+(i))+(1-w) B(Q*(i)) [0102] The interpolation parameter w is chosen such that the following cost function C(i)* is minimized*

[0102] C ( i ) = max { 1 N h = 1 H ( i ) { ( R ( P rec ( i ) ) - R ( + ( i , h ) ) ) 2 + ( G ( P rec ( i ) ) - G ( + ( i , h ) ) ) 2 + ( B ( P rec ( i ) ) - B ( + ( i , h ) ) ) 2 } , 1 N rec { ( R ( P rec ( i ) ) - R ( Q * ( i ) ) ) 2 + ( G ( P rec ( i ) ) - G ( Q * ( i ) ) ) 2 + ( B ( P rec ( i ) ) - B ( Q * ( i ) ) ) 2 } } ##EQU00005## [0103] Note that by minimizing the cost C(i), the distortion measure as described below is minimized. [0104] Different search strategies may be used to find the parameter w [0105] Use the closed form solution described below. [0106] No search: use w=0.5. [0107] Full search: choose a discrete set of values (w.sub.i).sub.i=1 … w in the interval [0,1] and evaluate C(i) for these values in order to find the w*, which minimizes C(i). [0108] Gradient descent search: start with w=0.5. Evaluate E1(i), E2(i) and C(i). Store C(i) and w as the lowest cost and its associated interpolation parameter w. If E1(i)>E2(i), update w based on the gradient of E1(i), else use the gradient of E2(i). Re-evaluate E1(i), E2(i), and C(i) at the new value of w. Compare the new cost C(i) to the lowest cost found so far. If new cost is higher than the lowest cost stop, else update the lowest cost and the associated value of w, and continue the gradient descent, where R(P.sub.rec(i)), G(P.sub.rec(i)), and B(P.sub.rec(i)) are the three unknowns to be determined.

[0109] In some embodiments, the above process could be performed with other color spaces and not necessarily the RGB color space. For example, the CIE 1931 XYZ or xyY, CIE 1976 Yu’v’, YCbCr, IPT, ICtCp, La*b*, or some other color model could be used instead. Furthermore, different weighting of the distortion of each component could be considered. Weighting based on illumination could also be considered, e.g. weighting distortion in dark areas more than distortion in bright areas. Other types of distortion, that include neighborhood information, could also be considered. That is, visibility of errors in a more sparse area is likely to be higher than in a more dense region, depending on the intensity of the current and neighboring samples. Such information could be considered in how the optimization is performed.

[0110] Down sampling and up sampling of chroma information may also consider geometry information, if available. That is, instead of down sampling and up sampling chroma information without consideration to geometry, the shape and characteristics of the point cloud around the neighborhood of the projected sample may be considered, and appropriately consider or exclude neighboring samples during these processes. In particular, neighboring samples for down sampling or interpolating may be considered that have a normal that is as similar as possible to the normal of the current sample. Weighting during filtering according to the normal difference as well as distance to the point may also be considered. This may help improve the performance of the down sampling and up sampling processes.

[0111] It should be noted that for some systems, up sampling of the Cb/Cr information may have to go through existing architectures, e.g. an existing color format converter, and it might not be possible to perform such guided up sampling. In those cases, only considerations for down sampling may be possible.

[0112] In some embodiments, it may be possible to indicate in the bit stream syntax the preferred method for up sampling the chroma information. A decoder (included in an encoder), in such a case, may try a variety of up sampling filters or methods, find the best performing one and indicate that in the bit stream syntax. On the decoder side, the decoder may know which up sampling method would perform best for reconstructing the full resolution YCbCr and consequently RGB data. Such method could be mandatory, but could also be optional in some architectures.

[0113] Clipping as well as other considerations for color conversion, may also apply to point cloud data and may be considered to further improve the performance of the point cloud compression system. Such methods may also apply to other color representations and not necessarily YCbCr data, such as the YCoCg and ICtCp representation. For such representations different optimization may be required due to the nature of the color transform.

Example Objective Evaluation Method

[0114] A point cloud consists of a set of points represented by (x,y,z) and various attributes of which color components (y,u,v) are of importance. First, define the point v. It has as a mandatory position in a 3D space (x,y,z) and an optional color attribute c that has components r,g,b or y,u,v and optional other attributes possibly representing normal or texture mappings.

point v=(((x, y, z), [c], [a.sub.0 … a.sub.A]): x,y,z.di-elect cons.R, [c.di-elect cons.(r,g,b)|r,g,b.di-elect cons.N], [a.sub.i.di-elect cons.[0,1]]) (def. 1)

[0115] The point cloud is then a set of K points without a strict ordering:

Original Point Cloud V.sub.or={(v.sub.i): i=0 … K-1} (def. 2)

[0116] The point cloud comprises a set of (x,y,z) coordinates and attributes that can be attached to the points. The original point cloud Vor will act as the reference for determining the quality of a second degraded point cloud Vdeg. Vdeg consists of N points, where N does not necessarily=K. Vdeg is a version of the point cloud with a lower quality possibly resulting from lossy encoding and decoding of Vor. This can result in a different point count N.

Degraded Point Cloud V.sub.deg={(v.sub.i): i=0 … N-1} (def. 3)

[0117] The quality metric Q_(point cloud) is computed from Vor and Vdeg and used for assessment as shown in FIG. 2F.

[0118] Table 1, below, outlines the metrics used for the assessment of the quality of a point cloud, in some embodiments. The geometric distortion metrics are similar as ones used for meshes based on haussdorf (Linf) and root mean square (L2), instead of distance to surface. This approach takes the distance to the closest/most nearby point in the point cloud (see definitions 4, 5, 6, and 7) into account. Peak signal to noise ratio (PSNR) is defined as the peak signal of the geometry over the symmetric Root Mean Square (RMS/rms) distortion (def 8.). For colors, a similar metric is defined; the color of the original cloud is compared to the most nearby color in the degraded cloud and peak signal to noise ratio (PSNR) is computed per YUV/YCbCr component in the YUV color space (def. 10). An advantage of this metric is that it corresponds to peak signal to noise ratio (PSNR) in Video Coding. The quality metric is supported in the 3DG PCC software.

TABLE-US-00001 TABLE 1 Assessment criteria for assessment of the point cloud quality of Vdeg, Q.sub.point_cloud d_symmetric_rms Symmetric rms distance between the point clouds (def. 5.) d_symmetric_haussdorf Symmetric haussdorf distance between the clouds (def. 7.) psnr_geom Peak signal to noise ratio geometry (vertex positions) (def. 8.) psnr_y Peak signal to noise ratio geometry (colors Y) (def. 10) psnr_u Peak signal to noise ratio geometry (colors U) (as def. 10 rep. y for u) psnr_v Peak signal to noise ratio geometry (colors V) (as def. 10 rep. y for v) d rms ( V or , V deg ) = 1 K .SIGMA. vo .di-elect cons. Vor vo - vd_nearest _neighbour 2 ##EQU00006## (def. 4) d.sub.symmetric_rms(V.sub.or,V.sub.deg) = max(d.sub.rms(V.sub.or, V.sub.deg), d.sub.rms (V.sub.deg, V.sub.or)) (def. 5) d.sub.haussdorf(V.sub.or, V.sub.deg) = max.sub.v.sub.o.sub..di-elect cons.V.sub.or,(||v.sub.o - v.sub.d_nearest _neighbour||.sub.2, v.sub.d is the point in (def. 6) Vdeg closest to v.sub.o (L2)) d.sub.symmetric_haussdorf(V.sub.or, V.sub.deg) = max(d.sub.haussdorf(V.sub.or, V.sub.deg), d.sub.haussdorf(V.sub.deg,V.sub.or) (def. 7) BBwidth = max((xmax - xmin), (ymax - ymin), (zmax - zmin) (def. 8) psnr.sub.geom = 10log.sub.10 (|BBwidth||.sub.2.sup.2/(d.sub.symmetric rms(V)).sup.2) (def. 9) d y ( V or , V deg ) = 1 K .SIGMA. vo .di-elect cons. Vor y ( vo ) - y ( v dnearest neighbour ) 2 ##EQU00007## (def. 10) psnr.sub.y = 10log.sub.10(|255||.sup.2/(d.sub.y(V.sub.or, V.sub.deg).sup.2) (def. 11)

[0119] In some embodiments, additional metrics that define the performance of a codec are outlined below in Table 2.

TABLE-US-00002 TABLE 2 Additional Performance Metrics Compressed size Complete compressed mesh size In point count K, the number of vertices in Vor Out point count N, number of vertices in Vdeg Bytes_geometry_layer Number of bytes for encoding the vertex positions Bytes_color_layer (opt) Number of bytes for encoding the colour attributes Bytes_att_layer (opt) Number of bytes for encoding the other attributes Encoder time (opt) Encoder time in ms on commodity hardware (optional) Decoder time (opt) Decoder time in ms on commodity hardware (optional)

Example Closed Form Solution

[0120] For each point P.sub.rec(i) in the reconstructed point cloud, let Q*(i) be its nearest neighbor in the original point cloud. For each point P.sub.rec(i) in the reconstructed point cloud, let (Q.sup.+(i, h)).sub.h.di-elect cons.{1, … , H(i)} be the set of point in the original point cloud that share P.sub.rec(i) as their nearest neighbor in the reconstructed point cloud. Let .sup.+(i) be the centroid of (Q.sup.+(i, h)).sub.h.di-elect cons.{1, … , H(i)}.

If H=0, then C(P.sub.rec(i)=C(Q*(i))

[0121] Denote as R-G-B vector C(P) associated with a given point P. In order to compute the color for a given P.sub.rec(i), we have the following formulation:

argmin C ( P rec ( i ) ) max { 1 N rec C ( P rec ( i ) ) - C ( Q * ( i ) ) 2 , 1 N h = 1 H C ( P rec ( i ) ) - C ( Q + ( i , h ) ) 2 } ##EQU00008## Where max { 1 N rec C ( P rec ( i ) ) - C ( Q * ( i ) ) 2 , h = 1 H C ( P rec ( i ) ) - C ( + ( i ) ) + C ( + ( i ) ) - C ( Q + ( i , h ) ) 2 } = max { 1 N rec C ( P rec ( i ) ) - C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) - C ( + ( i ) ) 2 + 1 N h = 1 H C ( + ( i ) ) - C ( Q + ( i , h ) ) 2 + 2 N h = 1 H C ( P rec ( i ) ) - C ( + ( i ) ) , C ( + ( i ) ) - C ( Q + ( i , h ) ) } = max { 1 N rec C ( P rec ( i ) ) - C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) - C ( + ( i ) ) 2 + 1 N h = 1 H C ( + ( i ) ) - C ( Q + ( i , h ) ) 2 } ##EQU00008.2##

[0122] Now denote D.sup.2=.SIGMA..sub.h=1.sup.H.parallel.C(.sup.+(i))-C(Q.sup.+(i, h)).parallel..sup.2,* so that*

argmin C ( P rec ( i ) ) max { 1 N rec C ( P rec ( i ) ) - C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) - C ( + ( i ) ) 2 + D 2 N } . ##EQU00009##

[0123] Note: if H=1 then D.sup.2=0

[0124] Let C.sup.0(P.sub.rec(i)) be a solution of the previous minimization problem. It can be shown that C.sup.0(P.sub.rec(i) could be expressed as:

C.sup.0(P.sub.rec(i))=wC(Q*(i)+(1-w)C(.sup.+(i))

[0125] Furthermore, C.sup.0(P.sub.rec(i)) verifies:

1 N rec wC ( Q * ( i ) ) + ( 1 - w ) C ( + ( i ) ) - C ( Q * ( i ) ) 2 = H N wC ( Q * ( i ) ) + ( 1 - w ) C ( + ( i ) ) - C ( + ( i ) ) 2 + D 2 N ##EQU00010## ( 1 - w ) 2 C ( + ( i ) ) - C ( Q * ( i ) ) 2 = w 2 HN rec N C ( Q * ( i ) ) - C ( + ( i ) 2 + D 2 N rec N ##EQU00010.2## Let .delta. 2 = C ( Q * ( i ) ) - C ( + ( i ) 2 and r = N rec N ##EQU00010.3## if .delta. 2 = 0 , then C ( P rec ( i ) ) = C ( Q * ( i ) ) = C ( + ( i ) ( 1 - 2 ) 2 .delta. 2 = w 2 r H .delta. 2 + rD 2 .delta. 2 + w 2 .delta. 2 - 2 w .delta. 2 = w 2 r H .delta. 2 + r D 2 .delta. 2 ( 1 - rH ) w 2 - 2 .delta. 2 w + ( .delta. 2 + r D 2 ) = 0 ( rH - 1 ) w 2 + 2 w + ( .alpha. r - 1 ) = 0 With .alpha. = D 2 .delta. 2 if H = 1 , then w = 1 2 if H > 1 .DELTA. = 4 - 4 ( rH - 1 ) ( .alpha. r - 1 ) .DELTA. = 4 - 4 ( rH - 1 ) .alpha. r + 4 H - 4 .DELTA. = 4 ( H - ( rH - 1 ) .alpha. r ) If .DELTA. = 0 w = - 1 ( rH - 1 ) If .DELTA. > 0 w 1 = - 1 - ( H - ( Hr - 1 ) .alpha. r ) ( rH - 1 ) w 2 = - 1 + ( H - ( Hr - 1 ) .alpha. r ) ( rH - 1 ) ##EQU00010.4##

[0126] Where the cost C(i) is computed for both w1 and w2 and the value that leads to the minimum cost is retained as the final solution.

Segmentation Process

[0127] FIG. 3A illustrates an example segmentation process for determining patches for a point cloud, according to some embodiments. The segmentation process as described in FIG. 3A may be performed by a decomposition into patches module, such as decomposition into patches module 206. A segmentation process may decompose a point cloud into a minimum number of patches (e.g., a contiguous subset of the surface described by the point cloud), while making sure that the respective patches may be represented by a depth field with respect to a patch plane. This may be done without a significant loss of shape information.

[0128] In some embodiments, a segmentation process comprises: [0129] Letting point cloud PC be the input point cloud to be partitioned into patches and {P(0), P(1) … , P(N-1)} be the positions of points of point cloud PC. [0130] In some embodiments, a fixed set D={D(0), D(1), … , D(K-1)} of K 3D orientations is pre-defined. For instance, D may be chosen as follows D={(1.0, 0.0, 0.0), (0.0, 1.0, 0.0), (0.0, 0.0, 1.0), (-1.0, 0.0, 0.0), (0.0, -1.0, 0.0), (0.0, 0.0, -1.0)} [0131] In some embodiments, the normal vector to the surface at every point P(i) is estimated. Any suitable algorithm may be used to determine the normal vector to the surface. For instance, a technique could include fetching the set H of the “N” nearest points of P(i), and fitting a plane .PI.(i) to H(i) by using principal component analysis techniques. The normal to P(i) may be estimated by taking the normal .gradient.(i) to .PI.(i). Note that “N” may be a user-defined parameter or may be found by applying an optimization procedure. “N” may also be fixed or adaptive. The normal values may then be oriented consistently by using a minimum-spanning tree approach. [0132] Normal-based Segmentation: An initial segmentation S0 of the points of point cloud PC may be obtained by associating respective points with the direction D(k) which maximizes the score .gradient.(i)|D(k), where .|. is the canonical dot product of R3. Pseudo code is provided below.

TABLE-US-00003 [0132] for (i = 0; i < pointCount; ++i) { clusterIndex = 0; bestScore = .gradient.(i)|D(0) ; for(j = 1; j < K; ++j) { score = .gradient.(i)|D(j) ; if (score > bestScore) { bestScore = score; clusterIndex = j; } } partition[i] = clusterIndex; }

[0133] Iterative segmentation refinement: Note that segmentation SO associates respective points with the plane .PI.(i) that best preserves the geometry of its neighborhood (e.g. the neighborhood of the segment). In some circumstances, segmentation S0 may generate too many small connected components with irregular boundaries, which may result in poor compression performance. In order to avoid such issues, the following iterative segmentation refinement procedure may be applied: [0134] 1. An adjacency graph A may be built by associating a vertex V(i) to respective points P(i) of point cloud PC and by adding R edges {E(i,j(0)), … , EN(R-1)} connecting vertex V(i) to its nearest neighbors {V(j(0)), V(j(1)), … , V(j(R-1))}. More precisely, {V(j(0)), V(j(1)), … , V(j(R-1))} may be the vertices associated with the points {P(j(0)), P(j(1)), … , P(j(R-1))}, which may be the nearest neighbors of P(i). Note that R may be a user-defined parameter or may be found by applying an optimization procedure. It may also be fixed or adaptive. [0135] 2. At each iteration, the points of point cloud PC may be traversed and every vertex may be associated with the direction D(k)* that maximizes*

[0135] ( .gradient. ( i ) | D ( k ) + .lamda. R .zeta. ( i ) ) , ##EQU00011##

where |.zeta.(i)| is the number of the R-nearest neighbors of V(i) belonging to the same cluster and .lamda. is a parameter controlling the regularity of the produced patches. Note that the parameters .lamda. and R may be defined by the user or may be determined by applying an optimization procedure. They may also be fixed or adaptive. In some embodiments, a “user” as referred to herein may be an engineer who configured a point cloud compression technique as described herein to one or more applications. [0136] 3.* An example of pseudo code is provided below*

TABLE-US-00004 [0136] for(I = 0; l < iterationCount; ++l) { for(i = 0; i < pointCount; ++i) { clusterIndex = partition [i]; bestScore = 0.0; for(k = 0; k < K; ++k) { score = .gradient.(i)|D(k) ; for(j .di-elect cons. {j(0), j(1), … , j(R - 1)}) { if (k == partition[j]) { score += .lamda. R ; ##EQU00012## } } if (score > bestScore) { bestScore = score; clusterIndex = k; } } partition[i] = clusterIndex; } } *In some embodiments, the pseudo code shown above may further include an early termination step. For example, if a score that is a particular value is reached, or if a difference between a score that is reached and a best score only changes by a certain amount or less, the search could be terminated early. Also, the search could be terminated if after a certain number of iterations (l = m), the clusterindex does not change.

[0137] Patch segmentation: In some embodiments, the patch segmentation procedure further segments the clusters detected in the previous steps into patches, which may be represented with a depth field with respect to a projection plane. The approach proceeds as follows, according to some embodiments: [0138] 1. First, a cluster-based adjacency graph with a number of neighbors R’ is built, while considering as neighbors only the points that belong to the same cluster. Note that R’ may be different from the number of neighbors R used in the previous steps. [0139] 2. Next, the different connected components of the cluster-based adjacency graph are extracted. Only connected components with a number of points higher than a parameter a are considered. Let CC={CC(0), CC(1), … , CC(M-1)} be the set of the extracted connected components. [0140] 3. Respective connected component CC(m) inherits the orientation D(m) of the cluster it belongs to. The points of CC(m) are then projected on a projection plane having as normal the orientation D(m), while updating a depth map, which records for every pixel the depth of the nearest point to the projection plane. [0141] 4. An approximated version of CC(m), denoted C’(m), is then built by associating respective updated pixels of the depth map with a 3D point having the same depth. Let PC’ be the point cloud obtained by the union of reconstructed connected components {CC’(0), CC’(1), … , CC’(M-1)} [0142] 5. Note that the projection reconstruction process may be lossy and some points may be missing. In order, to detect such points, every point P(i) of point cloud PC may be checked to make sure it is within a distance lower than a parameter .delta. from a point of PC’. If this is not the case, then P(i) may be marked as a missed point and added to a set of missed points denoted MP. [0143] 6. The steps 2-5 are then applied to the missed points MP. The process is repeated until MP is empty or CC is empty. Note that the parameters .delta. and a may be defined by the user or may be determined by applying an optimization procedure. They may also be fixed or adaptive. [0144] 7. A filtering procedure may be applied to the detected patches in order to make them better suited for compression. Example filter procedures may include: [0145] a. A smoothing filter based on the geometry/texture/attributes of the points of the patches (e.g., median filtering), which takes into account both spatial and temporal aspects. [0146] b. Discarding small and isolated patches. [0147] c. User-guided filtering. [0148] d. Other suitable smoothing filter techniques.

Layers

[0149] The image generation process described above consists of projecting the points belonging to each patch onto its associated projection plane to generate a patch image. This process could be generalized to handle the situation where multiple points are projected onto the same pixel as follows: [0150] Let H(u, v) be the set of points of the current patch that get projected to the same pixel (u,v). Note that H(u, v) may be empty, may have one point or multiple points. [0151] If H(u, v) is empty then the pixel is marked as unoccupied. [0152] If the H(u, v) has a single element, then the pixel is filled with the associated geometry/texture/attribute value. [0153] If H(u,v), has multiple elements, then different strategies are possible: [0154] Keep only the nearest point P0(u,v) for the pixel (u,v) [0155] Take the average or a linear combination of a group of points that are within a distance d from P0(u,v), where d is a user-defined parameter needed only the encoder side. [0156] Store two images: one for P0(u,v) and one to store the furthest point P1(u, v) of H(u, v) that is within a distance d from P0(u,v) [0157] Store N images containing a subset of H(u, v)

[0158] The generated patch images for point clouds with points at the same patch location, but different depths may be referred to as layers herein. In some embodiments, scaling/up-sampling/down-sampling could be applied to the produced patch images/layers in order to control the number of points in the reconstructed point cloud.

[0159] Guided up-sampling strategies may be performed on the layers that were down-sampled given the full resolution image from another “primary” layer that was not down-sampled.

[0160] Down-sampling could leverage the closed loop techniques as described below in regard to closed-loop color conversion, while exploiting a guided up-sampling strategy. For example, a generated layer may be encoded independently, which allows for parallel decoding and error resilience. Also encoding strategies, such as those specified by the scalable-HEVC standard, may be leveraged in order to support advanced functionalities such as spatial, SNR (signal to noise ratio), and color gamut scalability.

[0161] In some embodiments, a delta prediction between layers could be adaptively applied based on a rate-distortion optimization. This choice may be explicitly signaled in the bit stream.

[0162] In some embodiments, the generated layers may be encoded with different precisions. The precision of each layer may be adaptively controlled by using a shift+scale or a more general linear or non-linear transformation.

[0163] In some embodiments, an encoder may make decisions on a scaling strategy and parameters, which are explicitly encoded in the bit stream. The decoder may read the information from the bit stream and apply the right scaling process with the parameters signaled by the encoder.

[0164] In some embodiments, a video encoding motion estimation process may be guided by providing a motion vector map to the video encoder indicating for each block of the image frame, a 2D search center or motion vector candidates for the refinement search. Such information, may be trivial to compute since the mapping between the 3D frames and the 2D image frames is available to the point cloud encoder and a coarse mapping between the 2D image frames could be computed by using a nearest neighbor search in 3D.

[0165] The video motion estimation/mode decision/intra-prediction could be accelerated/improved by providing a search center map, which may provide guidance on where to search and which modes to choose from for each N.times.N pixel block.

[0166] Hidden/non-displayed pictures could be used in codecs such as AV1 and HEVC. In particular, synthesized patches could be created and encoded (but not displayed) in order to improve prediction efficiency. This could be achieved by re-using a subset of the padded pixels to store synthesized patches.

[0167] The patch re-sampling (e.g., packing and patch segmentation) process described above exploits solely the geometry information. A more comprehensive approach may take into account the distortions in terms of geometry, texture, and other attributes and may improve the quality of the re-sampled point clouds.

[0168] Instead of first deriving the geometry image and optimizing the texture image given said geometry, a joint optimization of geometry and texture could be performed. For example, the geometry patches could be selected in a manner that results in minimum distortion for both geometry and texture. This could be done by immediately associating each possible geometry patch with its corresponding texture patch and computing their corresponding distortion information. Rate-distortion optimization could also be considered if the target compression ratio is known.

[0169] In some embodiments, a point cloud resampling process described above may additionally consider texture and attributes information, instead of relying only on geometry.

[0170] Also, a projection-based transformation that maps 3D points to 2D pixels could be generalized to support arbitrary 3D to 2D mapping as follows: [0171] Store the 3D to 2D transform parameters or the pixel coordinates associated with each point [0172] Store X, Y,* Z coordinates in the geometry images instead of or in addition to the depth information*

Packing

[0173] In some embodiments, depth maps associated with patches, also referred to herein as “depth patch images,” such as those described above, may be packed into a 2D image frame. For example, a packing module, such as packing module 208, may pack depth patch images generated by a spatial image generation module, such as spatial image generation module 210. The depth maps, or depth patch images, may be packed such that (A) no non-overlapping block of T.times.T pixels contains depth information from two different patches and such that (B) a size of the generated image frame is minimized.

[0174] In some embodiments, packing comprises the following steps: [0175] The patches are sorted by height and then by width. The patches are then inserted in image frame (I) one after the other in that order. At each step, the pixels of image frame (I) are traversed in raster order, while checking if the current patch could be inserted under the two conditions (A) and (B) described above. If it is not possible then the height of (I) is doubled. [0176] This process is iterated until all the patches are inserted.

[0177] In some embodiments, the packing process described above may be applied to pack a subset of the patches inside multiples tiles of an image frame or multiple image frames. This may allow patches with similar/close orientations based on visibility according to the rendering camera position to be stored in the same image frame/tile, to enable view-dependent streaming and/or decoding. This may also allow parallel encoding/decoding.

[0178] In some embodiments, the packing process can be considered a bin-packing problem and a first decreasing strategy as described above may be applied to solve the bin-packing problem. In other embodiments, other methods such as the modified first fit decreasing (MFFD) strategy may be applied in the packing process.

[0179] In some embodiments, if temporal prediction is used, such as described for inter compression encoder 250, such an optimization may be performed with temporal prediction/encoding in addition to spatial prediction/encoding. Such consideration may be made for the entire video sequence or per group of pictures (GOP). In the latter case additional constraints may be specified. For example, a constraint may be that the resolution of the image frames should not exceed a threshold amount. In some embodiments, additional temporal constraints may be imposed, even if temporal prediction is not used, for example such as that a patch corresponding to a particular object view is not moved more than x number of pixels from previous instantiations.

[0180] FIG. 3B illustrates an example image frame comprising packed patch images and padded portions, according to some embodiments. Image frame 300 includes patch images 302 packed into image frame 300 and also includes padding 304 in space of image frame 300 not occupied by patch images. In some embodiments, padding, such as padding 304, may be determined so as to minimize incongruences between a patch image and the padding. For example, in some embodiments, padding may construct new pixel blocks that are replicas of, or are to some degree similar to, pixel blocks that are on the edges of patch images. Because an image and/or video encoder may encode based on differences between adjacent pixels, such an approach may reduce the number of bytes required to encode an image frame comprising of patch images and padding, in some embodiments.

[0181] In some embodiments, the patch information may be stored in the same order as the order used during the packing, which makes it possible to handle overlapping 2D bounding boxes of patches. Thus a decoder receiving the patch information can extract patch images from the image frame in the same order in which the patch images were packed into the image frame. Also, because the order is known by the decoder, the decoder can resolve patch image bounding boxes that overlap.

[0182] FIG. 3C illustrates an example image frame 312 with overlapping patches, according to some embodiments. FIG. 3C shows an example with two patches (patch image 1 and patch image 2) having overlapping 2D bounding boxes 314 and 316 that overlap at area 318. In order to determine to which patch the T.times.T blocks in the area 318 belong, the order of the patches may be considered. For example, the T.times.T block 314 may belong to the last decoded patch. This may be because in the case of an overlapping patch, a later placed patch is placed such that it overlaps with a previously placed patch. By knowing the placement order it can be resolved that areas of overlapping bounding boxes go with the latest placed patch. In some embodiments, the patch information is predicted and encoded (e.g., with an entropy/arithmetic encoder). Also, in some embodiments, U0, V0, DU0 and DV0 are encoded as multiples of T, where T is the block size used during the padding phase.

[0183] FIG. 3C also illustrates blocks of an image frame 312, wherein the blocks may be further divided into sub-blocks. For example block A1, B1, C1, A2, etc. may be divided into multiple sub-blocks, and, in some embodiments, the sub-blocks may be further divided into smaller blocks. In some embodiments, a video compression module of an encoder, such as video compression module 218 or video compression module 264, may determine whether a block comprises active pixels, non-active pixels, or a mix of active and non-active pixels. The video compression module may budget fewer resources to compress blocks comprising non-active pixels than an amount of resources that are budgeted for encoding blocks comprising active pixels. In some embodiments, active pixels may be pixels that include data for a patch image and non-active pixels may be pixels that include padding. In some embodiments, a video compression module may sub-divide blocks comprising both active and non-active pixels, and budget resources based on whether sub-blocks of the blocks comprise active or non-active pixels. For example, blocks A1, B1, C1, A2 may comprise non-active pixels. As another example block E3 may comprise active pixels, and block B6, as an example, may include a mix of active and non-active pixels.

[0184] In some embodiments, a patch image may be determined based on projections, such as projecting a point cloud onto a cube, cylinder, sphere, etc. In some embodiments, a patch image may comprise a projection that occupies a full image frame without padding. For example, in a cubic projection each of the six cubic faces may be a patch image that occupies a full image frame.

[0185] For example, FIG. 3D illustrates a point cloud being projected onto multiple projections, according to some embodiments.

[0186] In some embodiments, a representation of a point cloud is encoded using multiple projections. For example, instead of determining patches for a segment of the point cloud projected on a plane perpendicular to a normal to the segment, the point cloud may be projected onto multiple arbitrary planes or surfaces. For example, a point cloud may be projected onto the sides of a cube, cylinder, sphere, etc. Also multiple projections intersecting a point cloud may be used. In some embodiments, the projections may be encoded using conventional video compression methods, such as via a video compression module 218 or video compression module 264. In particular, the point cloud representation may be first projected onto a shape, such as a cube, and the different projections/faces projected onto that shape (i.e. front (320), back (322), top (324), bottom (326), left (328), right (330)) may all be packed onto a single image frame or multiple image frames. This information, as well as depth information may be encoded separately or with coding tools such as the ones provided in the 3D extension of the HEVC (3D-HEVC) standard. The information may provide a representation of the point cloud since the projection images can provide the (x,y) geometry coordinates of all projected points of the point cloud. Additionally, depth information that provides the z coordinates may be encoded. In some embodiments, the depth information may be determined by comparing different ones of the projections, slicing through the point cloud at different depths. When projecting a point cloud onto a cube, the projections might not cover all point cloud points, e.g. due to occlusions. Therefore additional information may be encoded to provide for these missing points and updates may be provided for the missing points.

[0187] In some embodiments, adjustments to a cubic projection can be performed that further improve upon such projections. For example, adjustments may be applied at the encoder only (non-normative) or applied to both the encoder and the decoder (normative).

[0188] More specifically, in some embodiments alternative projections may be used. For example, instead of using a cubic projection, a cylindrical or spherical type of a projection method may be used. Such methods may reduce, if not eliminate, redundancies that may exist in the cubic projection and reduce the number or the effect of “seams” that may exist in cubic projections. Such seams may create artifacts at object boundaries, for example. Eliminating or reducing the number or effect of such seams may result in improved compression/subjective quality as compared to cubic projection methods. For a spherical projection case, a variety of sub-projections may be used, such as the equirectangular, equiangular, and authagraph projection among others. These projections may permit the projection of a sphere onto a 2D plane. In some embodiments, the effects of seams may be de-emphasized by overlapping projections, wherein multiple projections are made of a point cloud, and the projections overlap with one another at the edges, such that there is overlapping information at the seams. A blending effect could be employed at the overlapping seams to reduce the effects of the seams, thus making them less visible.

[0189] In addition to, or instead of, considering a different projection method (such as cylindrical or spherical projections), in some embodiments multiple parallel projections may be used. The multiple parallel projections may provide additional information and may reduce a number of occluded points. The projections may be known at the decoder or signaled to the decoder. Such projections may be defined on planes or surfaces that are at different distances from a point cloud object. Also, in some embodiments the projections may be of different shapes, and may also overlap or cross through the point cloud object itself. These projections may permit capturing some characteristics of a point cloud object that may have been occluded through a single projection method or a patch segmentation method as described above.

[0190] For example, FIG. 3E illustrates a point cloud being projected onto multiple parallel projections, according to some embodiments. Point cloud 350 which includes points representing a coffee mug is projected onto parallel horizontal projections 352 that comprise planes orthogonal to the Z axis. Point cloud 350 is also projected onto vertical projections 354 that comprise planes orthogonal to the X axis, and is projected onto vertical projections 356 that comprise planes orthogonal to the Y axis. In some embodiments, instead of planes, multiple projections may comprise projections having other shapes, such as multiple cylinders or spheres.

Generating Images Having Depth

[0191] In some embodiments, only a subset of the pixels of an image frame will be occupied and may correspond to a subset of 3D points of a point cloud. Mapping of patch images may be used to generate geometry, texture and attribute images, by storing for each occupied pixel the depth/texture/attribute value of its associated point.

[0192] In some embodiments, spatial information may be stored with various variations, for example spatial information may: [0193] Store depth as a monochrome image [0194] Store depth as Y and keep U and V empty (where YUV is a color space, also RGB color space may be used). [0195] Store depth information for different patches in different color planes Y, U and V, in order to avoid inter-patch contamination during compression and/or improve compression efficiency (e.g., have correlated patches in the same color plane). Also, hardware codec capabilities may be utilized, which may spend the same encoding\decoding time independently of the content of the frame. [0196] Store depth patch images on multiple images or tiles that could be encoded and decoded in parallel. One advantage is to store depth patch images with similar/close orientations or based on visibility according to the rendering camera position in the same image/tile, to enable view-dependent streaming and/or decoding. [0197] Store depth as Y and store a redundant version of depth in U and V [0198] Store X, Y, Z coordinates in Y, U, and V [0199] Different bit depth (e.g., 8, 10 or 12-bit) and sampling (e.g., 420, 422, 444 … ) may be used. Note that different bit depth may be used for the different color planes.

Padding

[0200] In some embodiments, padding may be performed to fill the non-occupied pixels with values such that the resulting image is suited for video/image compression. For example, image frame padding module 216 or image padding module 262 may perform padding as described below.

[0201] In some embodiments, padding is applied on pixels blocks, while favoring the intra-prediction modes used by existing video codecs. More precisely, for each block of size B.times.B to be padded, the intra prediction modes available at the video encoder side are assessed and the one that produces the lowest prediction errors on the occupied pixels is retained. This may take advantage of the fact that video/image codecs commonly operate on pixel blocks with pre-defined sizes (e.g., 64.times.64, 32.times.32, 16.times.16 … ). In some embodiments, other padding techniques may include linear extrapolation, in-painting techniques, or other suitable techniques.

Video Compression

[0202] In some embodiments, a video compression module, such as video compression module 218 or video compression module 264, may perform video compression as described below.

[0203] In some embodiments, a video encoder may leverage an occupancy map, which describes for each pixel of an image whether it stores information belonging to the point cloud or padded pixels. In some embodiments, such information may permit enabling various features adaptively, such as de-blocking, adaptive loop filtering (ALF), or shape adaptive offset (SAO) filtering. Also, such information may allow a rate control module to adapt and assign different, e.g. lower, quantization parameters (QPs), and in an essence a different amount of bits, to the blocks containing the occupancy map edges. Coding parameters, such as lagrangian multipliers, quantization thresholding, quantization matrices, etc. may also be adjusted according to the characteristics of the point cloud projected blocks. In some embodiments, such information may also enable rate distortion optimization (RDO) and rate control/allocation to leverage the occupancy map to consider distortions based on non-padded pixels. In a more general form, weighting of distortion may be based on the “importance” of each pixel to the point cloud geometry. Importance may be based on a variety of aspects, e.g. on proximity to other point cloud samples, directionality/orientation/position of the samples, etc. Facing forward samples, for example, may receive a higher weighting in the distortion computation than backward facing samples. Distortion may be computed using metrics such as Mean Square or Absolute Error, but different distortion metrics may also be considered, such as SSIM, VQM, VDP, Hausdorff distance, and others.

Occupancy Map Compression

[0204] In some embodiments, an occupancy map compression module, such as occupancy map compression module 220, may compress an occupancy map as described below.

Example Occupancy Map Compression Techniques

[0205] In some embodiments, an occupancy map may be encoded in a hierarchical mode. Such a process may comprise: [0206] 1. A binary information for each B1.times.B2 pixel block (e.g., a rectangle that covers the entire image, or smaller blocks of different sizes such as 64.times.64, 64.times.32, 32.times.32 block, etc.) being encoded indicating whether the block is empty (e.g., has only padded pixels) or non-empty (e.g., has non-padded pixels). [0207] 2. If the block is non-empty, then a second binary information may be encoded to indicate whether the block is full (e.g., all the pixels are non-padded) or not. [0208] 3. The non-empty and non-full blocks may then be refined by considering their (B1/2).times.(B2/2) sub-blocks. [0209] 4. The steps 1-3 may be repeated until the size of the block reaches a certain block size B3.times.B4 (e.g., of size 4.times.4). At this level only the empty/non-empty information may be encoded. [0210] 5. An entropy-based codec may be used to encode the binary information in steps 1 and 2. For instance, context adaptive binary arithmetic encoders may be used. [0211] 6. The reconstructed geometry image may be leveraged to better encode the occupancy map. More precisely, the residual prediction errors may be used to predict whether a block is empty or not or full or not. Such an information may be incorporated by using a different context based on the predicted case or simply by encoding the binary value XORed with the predicted value.

[0212] In some embodiments, mesh-based codecs may be an alternative to the approach described above.

Additional Example Occupancy Map Compression Technique

[0213] In some embodiments, auxiliary information and the patch encoding order may be leveraged in order to efficiently compress a mapping information indicating for each T.times.T block (e.g., 16.times.16 block) to which patch it belongs to. This mapping may be explicitly encoded in the bit stream as follows: [0214] A list of candidate patches is created for each T.times.T block by considering all the patches that overlap with that block. [0215] The list of candidates is sorted in the reverse order of the patches. [0216] For each block, the index of the patch in this list is encoded by using an arithmetic or other form of an entropy encoder (e.g. UVLC or Huffman based). [0217] Note that empty blocks are assigned a special index, such as zero. [0218] The mapping information described above makes it possible to detect empty T.times.T blocks (e.g., blocks that contain only padded pixels). The occupancy information is encoded only for the non-empty T.times.T blocks (e.g., the blocks that contain at least one non-padded pixel). [0219] The occupancy map is encoded with a precision of a B0.times.B0 blocks. In order to achieve lossless encoding B0 is chosen to be 1. In some embodiments B0=2 or B0=4, which may result in visually acceptable results, while significantly reducing the number of bits required to encode the occupancy map. [0220] Binary values are associated with B0.times.B0 sub-blocks belonging to the same T.times.T block. Different strategies are possible. For instance, one could associate a value of 1 if the sub-block contains at least some non-padded pixels and 0 otherwise. If a sub-block has a value of 1 it is said to be full, otherwise it is an empty sub-block. [0221] If all the sub-blocks of a T.times.T block are full (e.g., have value 1). The block is said to be full. Otherwise, the block is said to be non-full. [0222] A binary information is encoded for each T.times.T block to indicate whether it is full or not. Various encoding strategies could be used. For instance, a context adaptive binary arithmetic encoder could be used. [0223] If the block is non-full, an extra information is encoded indicating the location of the full/empty sub-blocks. More precisely, the process may proceed as follows: [0224] Different traversal orders are defined for the sub-blocks. FIG. 12B, shows some examples. The traversal orders are predetermined and known to both the encoder and decoder. [0225] The encoder chooses one of the traversal orders and explicitly signals its index in the bit stream. [0226] The binary values associated with the sub-blocks are encoded by using a run-length encoding strategy. [0227] The binary value of the initial sub-block is encoded. Various encoding strategies could be used. For instance, fixed length coding or a context adaptive binary arithmetic encoder could be used. [0228] Continuous runs of 0s and 1 s are detected, while following the traversal order selected by the encoder. [0229] The number of detected runs is encoded. Various encoding strategies could be used. For instance, fixed length coding or a context adaptive binary arithmetic encoder, or a universal variable length encoder (UVLC) could be used. [0230] The length of each run, except of the last one, is then encoded. Various encoding strategies could be used. For instance, fixed length coding, a context adaptive binary arithmetic encoder, or a universal variable length encoder could be used.

[0231] Note that the symbol probabilities used during the arithmetic encoding could be initialized by using values explicitly signaled in the bit stream by the encoder in order to improve compression efficiency. Such information could be signaled at frame, slice, row(s) of blocks, or block level, or using a non-fixed interval. In that case, a system may have the ability to signal the initialization interval, or the interval adaptation could be predefined between encoder and decoder. For example, the interval could start with one block, and then increment by one block afterwards (e.g. using an adaptation positions of {1, 2, 3 … N-1 … } blocks.

[0232] The choice of the traversal order may have a direct impact on the compression efficiency. Different strategies are possible. For instance, the encoder could choose the traversal order, which would result in the lowest number of bits or the lowest number of runs. In some embodiments, hierarchical sub-blocks with variable sizes may be used.

[0233] In some embodiments, temporal prediction may be used for encoding/compressing occupancy maps as follows:

a. The occupancy map of the current frame may be predicted from the occupancy map of a reference frame (e.g. through a difference process assuming zero motion). The prediction could be done at the frame level, but could also be done at a sub-block level, e.g. signal 1 bit whether a block will be predicted temporally, or the original map for a block will be used instead. b. Prediction could be enhanced by using motion compensation and by associating a motion vector with each T.times.T block. c. The values of the current block may be XOR-ed with the values of the block referenced by the motion vector or the co-located block. If no prediction is used, the current block may be coded as is. d. Motion vectors could be integer, integer multiples, or can have sub-pixel precision. e. The encoding strategy described above may be applied to the results. f. The motion vectors of the current block may be predicted based on the motion vectors of the previously encoded blocks. For example, a list of candidate predicted motion vectors may be computed based on the motion vectors of spatially and/or temporally neighboring blocks that have already been encoded. The index of the best candidate to be used as a predictor and the difference can be explicitly encoded in the bit stream. The process may be similar to the process used in codecs such as AVC and HEVC among others. A reduction in temporal candidates may be performed similar to what is done in HEVC to reduce memory requirements. The residual motion vector can then be encoded using a technique such as context adaptive arithmetic encoding or UVLC. g. A skip mode may also be supported to indicate that the predicted block matches exactly the reference block. In that case, no residual motion vector is needed. h. Different block sizes could be used instead of sticking with T.times.T blocks. i. The choice of the block size and the motion vectors could be achieved by minimizing the number of bits required to encode the occupancy map. j. The process could also consider multiple references.

[0234] In some embodiments, additional techniques for encoding/compression of an occupancy map may include: [0235] Using clues included in the video picture to help encode the occupancy map, such as: [0236] Use high quantization parameters QPs (e.g., 51) or use skip mode for blocks composed of padded pixels only. [0237] The arithmetic encoding contexts could be adaptively adjusted based on information extracted from the video bit streams associated with the texture/geometry/motion frames. [0238] Group the binary values associated with pixels into 8-bit or 10-bit words and encode them with dictionary-based approaches such as the DEFLATE algorithm. [0239] Pixels could be grouped 4.times.2/5.times.2 blocks or by leveraging a zig zag scan. [0240] Only the pixels belonging to non-empty T.times.T blocks may be encoded. [0241] The mapping information indicating for each T.times.T block to which patch it belongs may encoded.

Auxiliary Patch-Information Compression

[0242] In some embodiments, for each patch, the following information may be encoded. For example, by auxiliary patch-info compression module 222. [0243] Its location (U0, V0) in the packed image frame and the extent of its 2D bounding box (DU0, DV0). [0244] Minimum/maximum/average/median depth value [0245] Index of the projection direction.

Point Cloud Resampling

[0246] In some embodiments, a point cloud resampling module, such as point cloud resampling module 252, may resample a point cloud as described below.

[0247] In some embodiments, dynamic point clouds may have a different number of points from one frame to another. Efficient temporal prediction may require mapping the points of the current frame, denoted CF, to the points of a reference frame, denoted RF. Signaling such a mapping in a bit stream may require a high number of bits and thus may be inefficient. Instead, re-sampling of a current frame CF may be performed so that the current frame CF has the same number of points as reference frame RF. More precisely, the points of reference frame RF may be displaced such that its shape matches the shape of current frame CF. As a second step, the color and attributes of current frame CF may be transferred to the deformed version of reference frame RF. The obtained frame CF’ may be considered as the re-sampled version of the current frame. The decision to compress the approximation CF’ of CF may be made by comparing the rate distortion costs of both options (e.g., encoding CF’ as inter-frame vs. encoding CF as intra-frame). In some embodiments, pre-adjusting RF may be performed in an effort to make it a better reference for future CF images. Resampling may comprise the following: [0248] First, normals of the points associated with current frame CF and reference frame RF may be estimated and oriented consistently. For every point P belonging to current frame CF (resp. Q belonging to RF), let .alpha.(P) (resp., .alpha.(Q)) be its position and .gradient.(P) (resp., .gradient.(Q)) its normal. A 6D vector, denoted .nu.(P) (resp., .nu.(Q)) is then associated with every point by combining its position and a weighted version of its normal in the same vector.

[0248] .upsilon. ( P ) = [ .alpha. ( P ) .gradient. ( P ) ] .upsilon. ( Q ) = [ .alpha. ( Q ) .gradient. ( Q ) ] , ##EQU00013##

where .epsilon. is a parameter controlling the importance of normal for positions. .epsilon. could be defined by the user or could be determined by applying an optimization procedure. They could also be fixed of adaptive. [0249] Two mappings from reference frame RF to current frame CF and from current frame CF to reference frame RF are computed as follows: [0250] Every point Q of reference frame RF is mapped to the point P(Q) of current frame CF that has the minimum distance to Q in the 6D space defined in the previous step. [0251] Every point P of current frame CF is mapped to the point Q(P) of reference frame RF that has the minimum distance to P in the 6D space defined in the previous step. Let .rho.(Q) be the set of points of current frame CF that are mapped to the same point Q. [0252] At each iteration [0253] The positions of the points of reference frame RF are updated as follows:

[0253] .alpha. ’ ( Q ) = w .alpha. ( P ( Q ) ) + ( 1 - w ) .rho. ( Q ) P .di-elect cons. .rho. ( Q ) .alpha. ( P ) , ##EQU00014##

where |.rho.(Q)| is the number of elements of .rho.(Q). The parameter w could be defined by the user or could be determined by applying an optimization procedure. It could also be fixed or adaptive. [0254] The previous updated step results usually in an irregular repartition of the points. In order to overcome such limitations, a Laplacian-based smoothing procedure is applied. The idea is to update the positions of the points such that they stay as close as possible to {.alpha.’(Q)}, while favoring a repartition as close as possible to the original point repartition in reference frame RF. More precisely, the following sparse linear system may be solved:

[0254] { .alpha. * ( Q ) } = argmin { .alpha. ’ ( Q ) } { Q .di-elect cons. RF .alpha. ” ( Q ) - .alpha. ’ ( Q ) 2 + .gamma. Q .di-elect cons. RF .alpha. ” ( Q ) - 1 R Q ’ .di-elect cons. N ( Q ) .alpha. ” ( Q ’ ) - .alpha. ( Q ) - 1 R Q ’ .di-elect cons. N ( Q ) .alpha. ( Q ’ ) 2 } , ##EQU00015##

where N(Q) is the set of the R nearest neighbors of Q in reference frame RF. [0255] The mappings between the updated RF’ point cloud and current frame CF are then updated as follows [0256] 1. Every point Q of RF’ is mapped to the point P(Q) of current frame CF that has the minimum distance to Q in the 3D space of positions. [0257] 2. Every point P of current frame CF is mapped to the point Q(P) of reference frame RF that has the minimum distance to P in the 3D space of positions. Let .rho.(Q) be the set of points of current frame CF that are mapped to the same point Q. [0258] This process is iterated until a pre-defined number of iterations is reached or there is no further change. [0259] At this stage, the color and attribute information is transferred from current frame CF to RF’* by exploiting the following formula*

[0259] A ( Q ) = w ( A ) A ( P ( Q ) ) + ( 1 - w ( A ) ) .rho. ( Q ) P .di-elect cons. .rho. ( Q ) A ( P ) , ##EQU00016##

where A stands for the texture or attribute to be transferred, |.rho.(Q)| is the number of elements of .rho.(Q). The parameter w(A) could be defined by the user or could be determined by applying an optimization procedure. It could also be fixed of adaptive.

3D Motion Compensation

[0260] In some embodiments, the positions, attributes and texture information may be temporally predicted by taking the difference between the value at current resampled frame minus a corresponding value, e.g. motion compensated value, from the reference frame. These values may be fed to the image generation stage to be stored as images. For example, such techniques may be performed by 3D motion compensation and delta vector prediction module 254.

Smoothing Filter

[0261] In some embodiments, a smoothing filter of a decoder, such as smoothing filter 244 or smoothing filter 276 of decoder 230 or decoder 280, may perform smoothing as described below.

[0262] In some embodiments, a reconstructed point cloud may exhibit discontinuities at the patch boundaries, especially at very low bitrates. In order to alleviate such a problem, a smoothing filter may be applied to the reconstructed point cloud. Applying the smoothing filter may comprise: [0263] By exploiting the occupancy map, both the encoder and the decoder may be able to detect boundary points, which are defined as being points belonging to B0.times.B0 blocks encoded during the last iteration of the hierarchical occupancy map compression procedure described in previous sections above. [0264] The boundary points may have their positions/attribute/texture updated. More precisely, respective boundary points may be assigned a smoothed position based on its R nearest neighbors in the point cloud. The smoothed position may be the centroid/median of the nearest neighbors. Another option may comprise fitting a plane or any smooth surface the nearest neighbor and assigning as a smoothed position the projection of the point on that surface. The number of parameters and/or the smoothing strategy may be chosen by a user or determined by applying an optimization strategy. They may be fixed for all the points or chosen adaptively. These parameters may be signaled in the bit stream. [0265] In order to reduce the computational complexity of the smoothing stage, a subsampled version of the reconstructed point cloud may be considered when looking for the nearest neighbors. Such subsampled version could be efficiently derived by considering a subsampled version of the geometry image and the occupancy map.

[0266] FIG. 4A illustrates a process for compressing attribute and spatial information of a point cloud, according to some embodiments.

[0267] At 402, a point cloud is received by an encoder. The point cloud may be captured, for example by one or more sensors, or may be generated, for example in software.

[0268] At 404, compressed point cloud information is determined, using any of the techniques described herein or using one more combinations of the techniques described herein.

[0269] At 406, a compressed point cloud is encoded using the compressed point cloud information determined at 404. The point cloud may be compressed using any of the techniques described herein.

[0270] FIG. 4B illustrates a process for decompressing attribute and spatial information of a point cloud, according to some embodiments.

[0271] At 403 an encoded point cloud is received. The point cloud may have been encoded using any of the encoding techniques described herein, such as patch images packed into an image frame that is then encoded by a video encoder. In some embodiments, the encoded point cloud may comprise point cloud projections, such as projections onto a cube, cylinder, sphere, etc. that are then encoded via a video encoder.

[0272] At 405, spatial and attribute information for the encoded point cloud is determined. For example, a video decoder may be used to decode video encoded packed images or projects. Spatial information may then be determined based on the packed images or projections and combined to determine spatial information for points of the point cloud. For example, depth information for points of a patch may be matched with X and Y information for the points of the patch to determine spatial information for the points of the patch in 3D space. In a similar manner other attributes, included in patch images such as color attributes, texture attributes, etc. may be matched with corresponding points to determine attribute values for the points. Also, in the case of multiple projections, the same point may be identified in more than one of the projections to determine spatial information for the point in 3D space.

[0273] At 407, a decompressed point cloud may be provided to a recipient device or module.

[0274] FIG. 4C illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a point cloud, according to some embodiments.

[0275] At 410, patches are determined for portions of point cloud. For example patches may be determined as described above. At 425 patch information for the patches may be generated and at 426, may be encoded to be sent to a decoder. In some embodiments, encoded patch information may be separately encoded from one or more image frames comprising packed patch images.

[0276] At 411, a first patch (or next patch is selected). At 412 a color patch image is generated for the points of the point cloud included in the patch. At 414, one or more addition attribute images, such as a texture attribute image is generated for the points of the point cloud included in the patch.

[0277] At 413, spatial information images are generated for the points of the point cloud included in the patch. In some embodiments, to generate the spatial information images, the points of the point cloud are projected, at 415, onto a patch plane perpendicular to a normal vector normal to a surface of the point cloud at the patch location. At 417 a first spatial image is generated for the patch based on the points being projected on the patch plane at 415. In addition, depth information for the points of the patch relative to the patch plane is determined at 416, and at 418 a depth patch image is generated based on the depth information determined at 416.

[0278] At 419, it is determined whether there are additional patches for which patch images are to be generated. If so, the process reverts to 411 for the next patch. If not, at 420 the patch images for the patches are packed into one or more image frames. In some embodiments, patch images for respective patches may be packed before patch images are determined for other patches. At 421, an occupancy map is generated based on where the patch images were placed when being packed into the one or more image frames. At 424, the occupancy map is encoded.

[0279] At 422, spaces in the one or more image frames that are not occupied by patch images are padded.

[0280] At 423, the one or more image frames are video encoded, such as in accordance with a high efficiency video coding (HEVC) standard.

[0281] FIG. 4D illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a moving or changing point cloud, according to some embodiments.

[0282] At 430, point cloud information for a previously encoded point cloud is received wherein the point cloud information represents a subsequent version of the previously encoded point cloud. For example, the subsequent version may be a representation of the point cloud at a subsequent moment in time, wherein the point cloud is moving or changing as time progresses.

[0283] At 431, it is determined if any new patches need to be determined for the point cloud. For example, an object not currently in the previously encoded point cloud may have been added to the point cloud. For example, the point cloud may be a point cloud of a road and a ball may have entered into the road. If there is a need to add a new patch, the occupancy map is updated at 433 to include the new patch and encoded at 434. Also, at 432 patch images are generated for the new patch in similar manner as described in 412-414. The generated patch images are included in packing at 443.

[0284] At 435, a first or next patch of the patches generated for the reference (previous) point cloud is selected. At 436, the points of the patch are re-sampled as described herein. At 437 motion vectors for the points included in the selected patch between the reference point cloud and the current point cloud are determined. At 440 the motion vectors are used to generate a relative motion patch image. For example, in some embodiments, generating a relative motion patch image may comprise, encoding, at 441, vector motion in different directions using different image characteristics, as described herein. At 438 an updated color patch image is generated. In some embodiments, the updated color patch image may encode residual values indicating differences in colors of the points of the point cloud included in the patch between the reference point cloud and the current point cloud. In a similar manner, at 439, other attribute update patch images may be generated.

[0285] At 442, it is determined whether there are additional patches to be evaluated. If so, the process reverts to 435 for the next patch. If not, at 443 the patch images for the patches are packed into one or more image frames. In some embodiments, patch images for respective patches may be packed before patch images are determined for other patches.

[0286] At 444, spaces in the one or more image frames that are not occupied by patch images are padded.

[0287] At 445, the one or more image frames are video encoded, such as in accordance with a high efficiency video coding (HEVC) standard.

[0288] FIG. 4E illustrates a decoder receiving image frames comprising patch images, patch information, and an occupancy map, and generating a decompressed representation of a point cloud, according to some embodiments.

[0289] At 450, an occupancy map is received by a decoder, at 451 patch information is received by the decoder. In some embodiments the occupancy map and the patch information may be encoded and the decoder may decode the occupancy map and the patch information (not shown). At 452, the decoder receives one or more encoded video image frames. At 452 the decoder identifies patch images in the one or more encoded video image frames and at 454 the decoder decodes the encoded video image frames. In some embodiments, the decoder may utilize the occupancy map and the patch information to identify active and non-active portions of the one or more encoded video images and may adjust one or more decoded parameters used to decode the encoded video images based on whether portions, e.g. blocks, sub-blocks, pixels, etc. comprise active or non-active information.

[0290] At 455, the decoder determines spatial information and/or attribute information for the points of the respective patches and at 456 generates a decompressed representation of the point cloud encoded in the one or more encoded video images.

[0291] In some embodiments, active and non-active portions of an image frame may be indicated by a “mask.” For example, a mask may indicate a portion of an image that is a padding portion or may indicate non-active points of a point cloud, such as points that are hidden from view in one or more viewing angles.

[0292] In some embodiments, a “mask” may be encoded along with patch images or projections. In some embodiments, a “mask” may show “active/available” points and “non-active/non-available” points in space. In some embodiments, a mask may be independent from a texture and a depth patch image. In some embodiments, a mask may be combined with other information, such as a texture or depth patch image. For example, by indicating that certain values in a signal range correspond to active points, e.g. values above 16 and below 235 in an 8 bit image, and that other values correspond to non-active points, e.g. values below 16 or values above 235 in an 8 bit image. In some embodiments, additional considerations may be taken to avoid or reduce contamination between active and non-active regions. For example, it may be necessary to make use of lossless or visually lossless coding at the boundaries between active and non-active regions.

[0293] In some embodiments, a mask may be utilized in a variety of ways for improving coding efficiency. For example, a mask may be used with projection methods such as cylindrical, spherical or multiple projection as wells as decomposition into patches. In addition, a mask may be used with a cubic projection method.

[0294] FIG. 4F illustrates an encoder, adjusting encoding based on one or more masks for a point cloud, according to some embodiments.

[0295] At 462, an encoder receives a point cloud. At 464, the encoder generate multiple projections or patch images as described herein, for the received point cloud. At 466, the encoder determines or more masks. The masks may be hidden points, padded portions of an image frame, points not viewable from a particular view-point, etc. At 468, the encoder adjusts one or more encoding parameters based on the masks. For example the encoder may adjust a budget allocated to masked portions. Additional adjustments that an encoder may perform are described. At 468, the encoder encodes a compressed point cloud, for example via one or more video encoded image frames.

[0296] FIG. 4G illustrates a decoder, adjusting decoding based on one or more masks for a point cloud, according to some embodiments.

[0297] At 470, a decoder receives an encoded point cloud. At 472, the decoder determines one or more masks for portions of the encoded point cloud. For example the encoder may determine portions of image frames representing the compressed point cloud correspond to padding. Or, for a particular view of the point cloud being rendered by the decoder, the decoder may determine that one or more points of the compressed point cloud are not viewable from the particular point of view. In some embodiments, mask information may indicate which points are hidden when the point cloud is viewed from particular points of view. At 474, the decoder adjusts one or more decoding parameters based on the masks. Adjustments that may be made by a decoder based on active/non-active regions or points (e.g. masks) are described in more detail below. At 476 the decoder decodes the compressed point cloud.

[0298] In some embodiments, a mask may be used when performing motion estimation and mode decision. Commonly distortion is computed for an entire block. However, some blocks may have blocks that contain a combination of texture data as well as empty/nonvisible areas. For these areas only the textured data are of interest and any distortion in the non-visible areas may be ignored. Therefore, since commonly when performing such processes as motion estimation and mode decision, a distortion computation, such as Sum of Absolute Differences (SAD) or Sum of Square Errors (SSE), is performed, a mask may be used to alter the computation to exclude distortion for the non-visible areas. For example, for the SAD case, distortion may be computed by computing the sum of absolute differences of only samples in a block that correspond to a visible area in a current image. All other samples may be ignored during the computation. In some embodiments, distortion may be normalized at the pixel level thus avoiding having to consider blocks with different number of pixels.

[0299] In some embodiments, instead of only considering non-visible samples, samples that are adjacent to non-visible samples, or samples identified to correspond to different projections (but are placed when encoding within the same coding block) may be assigned different weights. For example samples in particular blocks could be considered more important for subjective quality, and a lower distortion tolerance may be assigned. In such case, the weighting for those samples may be increased, thus biasing decisions where the distortion for those samples is lower. Knowledge also that different samples in the same block of a particular size M.times.N during motion estimation or mode decision correspond to different objects, may also help with the determination of the block partitioning mode, e.g. the encoder could make an early decision (based potentially on a preliminary search) on whether different partitioning could/should be used.

[0300] In some embodiments, masks may be used for rate control and rate allocation. For example, it may be desirable that blocks that correspond to areas that contain both visible and non-visible samples be encoded at a different, and some times higher, quality than blocks that contain only visible samples. This is done so as to avoid leakage between visible and not visible samples and ensure the best quality at the point-clouds “boundaries”. Different quality may also be assigned based on depth information, which may also be available on the encoder. Flatter areas may tolerate much more distortion than areas with considerable variance in depth. Control of quality may be performed by adjusting quantization parameters/factors, but also by adjusting other parameters such as the lagrangian multiplier during mode decision, using different quantization matrices if available, enabling and/or adjusting quantization thresholding and the size and/or shapes of zonal quantization.

[0301] Quantization may also be adjusted according to the projection method used. If, for example an equirectangular projection method was used to project the object onto a sphere and then onto a 2D plane, it might be desirable to increase quantization on the top and bottom boundaries, and slowly decrease it when moving towards the center/equator. This may help compensate for some of the differences in resolution allocation when using a particular projection method. Different adjustments may also be made to the different color components, again based on similar assumptions, and in consideration again of the mask information.

[0302] Quantization may also be performed while considering whether a sample is a visible or a non-visible sample. For example, if a strategy involves the use of dynamic programming/trellis quantization methods for determining the value of a quantized coefficient. In such embodiments, an impact in distortion of a quantized coefficient, as well as its impact on bitrate at multiple reconstruction points may commonly be computed. This may be done for all coefficients while considering their bitrate interactions. Finally a decision may be made for all coefficients jointly by selecting the quantized values that would together result in the best rate distortion performance. In some embodiments, the visible and non-visible areas may be considered when computing such metrics.

[0303] Similar to the motion estimation and mode decision, processes, sample adaptive offset (SAO) techniques also commonly compute the resulting distortion for each possible mode or SAO value that may be used. Again, the decision may exclude non-visible samples, or prioritize, with different weights samples that are close to non-visible samples or samples that correspond to areas with considerably varying depth.

[0304] In some embodiments, masks may be used in any other coding process that may involve a distortion computation.

[0305] In some embodiments, masks may be used in preprocessing/prefiltering. For example, samples may be prefiltered based on their proximity to non-visible samples so as to reduce the possibility of artifacts and/or remove noise that may make encoding more difficult. Any form of prefiltering, including spatio-temporal filters, may be used.

[0306] In some embodiments, prefiltering may be applied to both texture as well as depth information.

[0307] Decisions in quantization parameters could also be made at the picture level (temporally) given the amount of visible/non-visible samples and depth variance on different pictures. Such decisions could be quite useful, for example, in a multi-pass coding system where analyze the entire sequence is first analyzed to determine the complexity and relationship of each frame with other frames. The coding parameters may then be decided that will be used for that frame in relationship to all other frames and given an expected quality or bitrate target. Similar decisions may also be made, not only for quantization parameters, but also for the picture coding types (i.e. I, P, or B), structures (e.g. hierarchical or not coding of N frames that follows a particular coding order of frames), references to use, weighting parameters, etc.

Encoding and Decoding (Normative Concepts)

[0308] Since a mask is likely to be available losslessly or visually losslessly at the decoder, as well as the depth information, this information may also be used at the decoder (and of course at the encoder) to further improve quality.

[0309] For example, deblocking and sample adaptive offset (SAO), as well as adaptive loop filtering (ALF) and deringing (in codecs that support such mechanisms), with exclusion of non-visible samples, samples that correspond to different projections, or samples with very different depth characteristics may use masking information. Instead, it may be desirable to only consider for such filtering methods samples that correspond to the same projection and are not so far from each other (depth wise). This may reduce blockiness and/or other artifacts that these methods try to mitigate. Other future types of in-loop post filtering may also be performed in a similar manner.

[0310] As another example, out of loop post filtering with visible/non-visible/different area segmentation may utilize masking information.

[0311] Implicit adjustment of QP quality parameters based on a certain percentage of visible/non-visible samples within a block may be performed. This may reduce signaling of coding parameters if such switching occurs frequently in a bit stream.

[0312] Adjustment of the transform type based on the percentage of visible/non-visible samples may be performed, including the consideration of shape adaptive discrete cosine transforms (DCT transforms).

[0313] Adjustment of overlapped block motion compensation techniques may utilize masking information, if existing in a codec, to mask away non-visible samples. A similar consideration may be performed for block motion compensation and/or intra prediction (including an intra block copy method). Samples that are considered visible may be considered when constructing a prediction signal, including also when interpolating to perform subpixel motion compensation or when performing bi-prediction. Masks from the current picture may be considered, but if desired, both the masks from the current picture as well as the masks corresponding to the reference pictures could be considered. Such considerations may be made adaptive at the encoder, through some form of signaling, i.e. at the sequence, picture, tile, slice, or even CTU/block level.

[0314] In some embodiments, clipping of the final value based on the mask or depth information may be performed.

[0315] In some embodiments, other prediction methods that may exist inside a codec (e.g. in AV1 or the Versatile Video Coding (VVC) standard currently being developed by the JVET team in MPEG) may be similarly adjusted or constrained based on the existence (and amount) of visible and non-visible points, and points corresponding to different projections.

[0316] In some embodiments, different control/adjustments may be applied to different color components as well as to the depth information.

Occupancy Map Compression

[0317] FIG. 4H illustrates more detail regarding compression of an occupancy map, according to some embodiments. The steps shown in FIG. 4H may be performed as part of steps 421 or 433 as described above. In some embodiments, any of the occupancy map compression techniques described herein may be performed at 421 or 433.

[0318] At 480 a list of candidate patches is determined for each block or modified block of an occupancy map.

[0319] At 481, the lists of candidate patches for each block are ordered in a reverse order as an order in which the patches were packed into the image frame. For example, the patches may be packed into an image, with larger patches packed before smaller patches. In contrast, the candidate list for each block of an occupancy map may include smaller patches before larger patches. At 482, an arithmetic encoder may be used to encode the patch candidate list for each block. In some embodiments, an entropy encoder may be used. Also, in some embodiments, empty blocks may be assigned a special value, such as zero, whereas patch candidates may be assigned numbers corresponding to a patch number, such as 1, 2, 3, etc.

[0320] At 483, for each block sub-blocks are determined according to a determined precision value. The determined precision value may be encoded with the occupancy map such that a decoder may determine the determined precision value used at the encoder.

[0321] At 484, for each block, a binary value (e.g. 0 or 1) is determined for each sub-block of the block. Full sub-blocks are assigned a different binary value than non-full sub-blocks. If all sub-blocks of a block are full, the block may be assigned a binary “full” value.

[0322] At 485, for each non-full and non-empty block, a traversal order is determined. For example, any of the example traversal orders shown in FIG. 4I (or other traversal orders) may be determined. For example, a traversal order may be diagonal from bottom left to top right as shown in example traversal 1, top-to-bottom as shown in example traversal 2 (or bottom-to-top), diagonal from top right to bottom left as shown in example traversal 3, or side-to-side, as shown in example traversal 4. A run-length encoding strategy as described above in regard to occupancy map compression may be used to encode the binary values for the sub-blocks using the determined traversal order.

[0323] FIG. 4I illustrates example blocks and traversal patterns for compressing an occupancy map, according to some embodiments. The traversal patterns may be used as described above in regard to occupancy map compression and in FIG. 4H.

Example Applications Using Point Cloud Encoders and Decoders

[0324] FIG. 5 illustrates compressed point clouds being used in a 3-D telepresence application, according to some embodiments.

[0325] In some embodiments, a sensor, such as sensor 102, an encoder, such as encoder 104 or any of the other encoders described herein, and a decoder, such as decoder 116 or any of the decoders described herein, may be used to communicate point clouds in a 3-D telepresence application. For example, a sensor, such as sensor 102, at 502 may capture a 3D image and at 504, the sensor or a processor associated with the sensor may perform a 3D reconstruction based on sensed data to generate a point cloud.

[0326] At 506, an encoder such as encoder 104 may compress the point cloud and at 508 the encoder or a post processor may packetize and transmit the compressed point cloud, via a network 510. At 512, the packets may be received at a destination location that includes a decoder, such as decoder 116. The decoder may decompress the point cloud at 514 and the decompressed point cloud may be rendered at 516. In some embodiments a 3-D telepresence application may transmit point cloud data in real time such that a display at 516 represents images being observed at 502. For example, a camera in a canyon may allow a remote user to experience walking through a virtual canyon at 516.

[0327] FIG. 6 illustrates compressed point clouds being used in a virtual reality (VR) or augmented reality (AR) application, according to some embodiments.

[0328] In some embodiments, point clouds may be generated in software (for example as opposed to being captured by a sensor). For example, at 602 virtual reality or augmented reality content is produced. The virtual reality or augmented reality content may include point cloud data and non-point cloud data. For example, a non-point cloud character may traverse a landscape represented by point clouds, as one example. At 604, the point cloud data may be compressed and at 606 the compressed point cloud data and non-point cloud data may be packetized and transmitted via a network 608. For example, the virtual reality or augmented reality content produced at 602 may be produced at a remote server and communicated to a VR or AR content consumer via network 608. At 610, the packets may be received and synchronized at the VR or AR consumer’s device. A decoder operating at the VR or AR consumer’s device may decompress the compressed point cloud at 612 and the point cloud and non-point cloud data may be rendered in real time, for example in a head mounted display of the VR or AR consumer’s device. In some embodiments, point cloud data may be generated, compressed, decompressed, and rendered responsive to the VR or AR consumer manipulating the head mounted display to look in different directions.

[0329] In some embodiments, point cloud compression as described herein may be used in various other applications, such as geographic information systems, sports replay broadcasting, museum displays, autonomous navigation, etc.

Example Computer System

[0330] FIG. 7 illustrates an example computer system 700 that may implement an encoder or decoder or any other ones of the components described herein, (e.g., any of the components described above with reference to FIGS. 1-6), in accordance with some embodiments. The computer system 700 may be configured to execute any or all of the embodiments described above. In different embodiments, computer system 700 may be any of various types of devices, including, but not limited to, a personal computer system, desktop computer, laptop, notebook, tablet, slate, pad, or netbook computer, mainframe computer system, handheld computer, workstation, network computer, a camera, a set top box, a mobile device, a consumer device, video game console, handheld video game device, application server, storage device, a television, a video recording device, a peripheral device such as a switch, modem, router, or in general any type of computing or electronic device.

[0331] Various embodiments of a point cloud encoder or decoder, as described herein may be executed in one or more computer systems 700, which may interact with various other devices. Note that any component, action, or functionality described above with respect to FIGS. 1-6 may be implemented on one or more computers configured as computer system 700 of FIG. 7, according to various embodiments. In the illustrated embodiment, computer system 700 includes one or more processors 710 coupled to a system memory 720 via an input/output (I/O) interface 730. Computer system 700 further includes a network interface 740 coupled to I/O interface 730, and one or more input/output devices 750, such as cursor control device 760, keyboard 770, and display(s) 780. In some cases, it is contemplated that embodiments may be implemented using a single instance of computer system 700, while in other embodiments multiple such systems, or multiple nodes making up computer system 700, may be configured to host different portions or instances of embodiments. For example, in one embodiment some elements may be implemented via one or more nodes of computer system 700 that are distinct from those nodes implementing other elements.

[0332] In various embodiments, computer system 700 may be a uniprocessor system including one processor 710, or a multiprocessor system including several processors 710 (e.g., two, four, eight, or another suitable number). Processors 710 may be any suitable processor capable of executing instructions. For example, in various embodiments processors 710 may be general-purpose or embedded processors implementing any of a variety of instruction set architectures (ISAs), such as the x86, PowerPC, SPARC, or MIPS ISAs, or any other suitable ISA. In multiprocessor systems, each of processors 710 may commonly, but not necessarily, implement the same ISA.

[0333] System memory 720 may be configured to store point cloud compression or point cloud decompression program instructions 722 and/or sensor data accessible by processor 710. In various embodiments, system memory 720 may be implemented using any suitable memory technology, such as static random access memory (SRAM), synchronous dynamic RAM (SDRAM), nonvolatile/Flash-type memory, or any other type of memory. In the illustrated embodiment, program instructions 722 may be configured to implement an image sensor control application incorporating any of the functionality described above. In some embodiments, program instructions and/or data may be received, sent or stored upon different types of computer-accessible media or on similar media separate from system memory 720 or computer system 700. While computer system 700 is described as implementing the functionality of functional blocks of previous Figures, any of the functionality described herein may be implemented via such a computer system.

[0334] In one embodiment, I/O interface 730 may be configured to coordinate I/O traffic between processor 710, system memory 720, and any peripheral devices in the device, including network interface 740 or other peripheral interfaces, such as input/output devices 750. In some embodiments, I/O interface 730 may perform any necessary protocol, timing or other data transformations to convert data signals from one component (e.g., system memory 720) into a format suitable for use by another component (e.g., processor 710). In some embodiments, I/O interface 730 may include support for devices attached through various types of peripheral buses, such as a variant of the Peripheral Component Interconnect (PCI) bus standard or the Universal Serial Bus (USB) standard, for example. In some embodiments, the function of I/O interface 730 may be split into two or more separate components, such as a north bridge and a south bridge, for example. Also, in some embodiments some or all of the functionality of I/O interface 730, such as an interface to system memory 720, may be incorporated directly into processor 710.

[0335] Network interface 740 may be configured to allow data to be exchanged between computer system 700 and other devices attached to a network 785 (e.g., carrier or agent devices) or between nodes of computer system 700. Network 785 may in various embodiments include one or more networks including but not limited to Local Area Networks (LANs) (e.g., an Ethernet or corporate network), Wide Area Networks (WANs) (e.g., the Internet), wireless data networks, some other electronic data network, or some combination thereof. In various embodiments, network interface 740 may support communication via wired or wireless general data networks, such as any suitable type of Ethernet network, for example; via telecommunications/telephony networks such as analog voice networks or digital fiber communications networks; via storage area networks such as Fibre Channel SANs, or via any other suitable type of network and/or protocol.

[0336] Input/output devices 750 may, in some embodiments, include one or more display terminals, keyboards, keypads, touchpads, scanning devices, voice or optical recognition devices, or any other devices suitable for entering or accessing data by one or more computer systems 700. Multiple input/output devices 750 may be present in computer system 700 or may be distributed on various nodes of computer system 700. In some embodiments, similar input/output devices may be separate from computer system 700 and may interact with one or more nodes of computer system 700 through a wired or wireless connection, such as over network interface 740.

[0337] As shown in FIG. 7, memory 720 may include program instructions 722, which may be processor-executable to implement any element or action described above. In one embodiment, the program instructions may implement the methods described above. In other embodiments, different elements and data may be included. Note that data may include any data or information described above.

[0338] Those skilled in the art will appreciate that computer system 700 is merely illustrative and is not intended to limit the scope of embodiments. In particular, the computer system and devices may include any combination of hardware or software that can perform the indicated functions, including computers, network devices, Internet appliances, PDAs, wireless phones, pagers, etc. Computer system 700 may also be connected to other devices that are not illustrated, or instead may operate as a stand-alone system. In addition, the functionality provided by the illustrated components may in some embodiments be combined in fewer components or distributed in additional components. Similarly, in some embodiments, the functionality of some of the illustrated components may not be provided and/or other additional functionality may be available.

[0339] Those skilled in the art will also appreciate that, while various items are illustrated as being stored in memory or on storage while being used, these items or portions of them may be transferred between memory and other storage devices for purposes of memory management and data integrity. Alternatively, in other embodiments some or all of the software components may execute in memory on another device and communicate with the illustrated computer system via inter-computer communication. Some or all of the system components or data structures may also be stored (e.g., as instructions or structured data) on a computer-accessible medium or a portable article to be read by an appropriate drive, various examples of which are described above. In some embodiments, instructions stored on a computer-accessible medium separate from computer system 700 may be transmitted to computer system 700 via transmission media or signals such as electrical, electromagnetic, or digital signals, conveyed via a communication medium such as a network and/or a wireless link. Various embodiments may further include receiving, sending or storing instructions and/or data implemented in accordance with the foregoing description upon a computer-accessible medium. Generally speaking, a computer-accessible medium may include a non-transitory, computer-readable storage medium or memory medium such as magnetic or optical media, e.g., disk or DVD/CD-ROM, volatile or non-volatile media such as RAM (e.g. SDRAM, DDR, RDRAM, SRAM, etc.), ROM, etc. In some embodiments, a computer-accessible medium may include transmission media or signals such as electrical, electromagnetic, or digital signals, conveyed via a communication medium such as network and/or a wireless link.

[0340] The methods described herein may be implemented in software, hardware, or a combination thereof, in different embodiments. In addition, the order of the blocks of the methods may be changed, and various elements may be added, reordered, combined, omitted, modified, etc. Various modifications and changes may be made as would be obvious to a person skilled in the art having the benefit of this disclosure. The various embodiments described herein are meant to be illustrative and not limiting. Many variations, modifications, additions, and improvements are possible. Accordingly, plural instances may be provided for components described herein as a single instance. Boundaries between various components, operations and data stores are somewhat arbitrary, and particular operations are illustrated in the context of specific illustrative configurations. Other allocations of functionality are envisioned and may fall within the scope of claims that follow. Finally, structures and functionality presented as discrete components in the example configurations may be implemented as a combined structure or component. These and other variations, modifications, additions, and improvements may fall within the scope of embodiments as defined in the claims that follow.