Method and apparatus for video encoding for machine vision

Abstract

A technique for encoding video for machine vision and human / machine hybrid vision, comprising: receiving an input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision; compressing the input using the first codec, the compressing including downsampling the input using a downsampling module and upsampling the compressed input using an upsampling module to generate a residual signal; quantizing the residual signal to obtain a quantized representation of the input; entropy encoding the quantized representation of the input using one or more convolutional filter modules; and training one or more networks using the entropy encoded quantized representation.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority under 35 U.S.C. § 119 to U.S. Provisional Patent Application No. 63 / 277,517, filed November 9, 2021, and U.S. Patent Application No. 17 / 950,564, filed September 22, 2022, the disclosures of which are incorporated herein by reference in their entireties.

[0002] This disclosure relates to video coding for machine vision. In particular, techniques for encoding video for machine vision and human / machine hybrid vision are disclosed. [Background technology]

[0003] Traditionally, videos or images are consumed by humans for a variety of uses, such as entertainment, education, etc. Therefore, video coding or image coding often exploits the properties of the human visual system for better compression efficiency while maintaining good subjective quality.

[0004] In recent years, with the increasing application of machine learning, many intelligent platforms utilize a multitude of sensors as well as video for machine vision tasks such as object detection, segmentation, and tracking. How to encode video or images for consumption by machine tasks has become an interesting and challenging problem, leading to the introduction of machine-oriented video coding (VCM) research. To this end, the international standards group MPEG has created an ad-hoc group, "Machine-oriented Video Coding (VCM)," to standardize related technologies for better interoperability between different devices.

[0005] Existing video codecs are primarily intended for human consumption. However, an increasing amount of video is being consumed by machines for machine vision tasks such as object detection, instance segmentation, and object tracking. It is important to develop video codecs that efficiently encode video for machine vision or hybrid machine / human vision. Summary of the Invention [Means for solving the problem]

[0006] The following presents a simplified summary of one or more embodiments of the present disclosure in order to provide a basic understanding of such embodiments. This summary is not an extensive overview of all possible embodiments, and is not intended to identify key or critical elements of all embodiments or to delineate the scope of any or all embodiments. Its sole purpose is to present some concepts of one or more embodiments of the present disclosure in a simplified form as a prelude to the more detailed description that is presented later.

[0007] A method, apparatus, and non-transitory computer-readable medium for encoding video for machine vision and human / machine hybrid vision.

[0008] According to one exemplary embodiment, a method for encoding video for machine vision and human / machine hybrid vision, executed by one or more processors, includes receiving an input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision. The method further includes compressing the input using the first codec, the compressing step including downsampling the input using a downsampling module and upsampling the compressed input using an upsampling module to generate a residual signal. The method further includes quantizing the residual signal to obtain a quantized representation of the input. The method further includes entropy coding the quantized representation of the input using one or more convolutional filter modules and training one or more networks using the entropy-coded quantized representation. According to one exemplary embodiment, an apparatus for encoding video for machine vision and human / machine hybrid vision includes at least one memory configured to store computer program code and at least one processor configured to access the computer program code and operate as instructed by the computer program code. The computer program code includes configuration code configured to cause the at least one processor to receive input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision. The computer program code further includes compression code configured to cause the at least one processor to compress the input using the first codec, the compression code including downsampling code configured to cause the at least one processor to downsample the input using a downsampling module, and the compression code including upsampling code configured to cause the at least one processor to upsample the compressed input using an upsampling module that generates a residual signal. The computer program code further includes quantization code configured to cause the at least one processor to quantize the residual signal to obtain a quantized representation of the input. The computer program code further includes entropy encoding code configured to cause the at least one processor to encode the quantized representation of the input using one or more convolutional filter modules. The computer program code further includes training code configured to train the one or more networks using the entropy encoded quantized representation. According to one exemplary embodiment, a non-transitory computer-readable medium having instructions stored thereon, when executed by a processor, causes the processor to perform a method for encoding video for machine vision and human-machine hybrid vision. The method includes receiving an input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision. The method further includes compressing the input using the first codec, the compressing step including downsampling the input using a downsampling module and upsampling the compressed input using an upsampling module to generate a residual signal. The method further includes quantizing the residual signal to obtain a quantized representation of the input. The method further includes entropy encoding the quantized representation of the input using one or more convolutional filter modules and training one or more networks using the entropy-encoded quantized representation.

[0009] Additional embodiments will be set forth in the description that follows, and in part will be obvious from the description and / or may be learned by practice of presented embodiments of the present disclosure.

[0010] These and other features and aspects of embodiments of the present disclosure will become apparent from the following description taken in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 2 is a diagram of an exemplary network device, according to various embodiments of the present disclosure. [Figure 2] FIG. 1 illustrates an architecture of the disclosed hybrid video codec, according to one embodiment of the present disclosure. [Figure 3]FIG. 1 illustrates video encoding for a machine system according to various embodiments of the present disclosure. [Figure 4] FIG. 1 illustrates an architecture of a learning-based image codec according to various embodiments of the present disclosure. [Figure 5] FIG. 1 illustrates an architecture of a learning-based image codec according to various embodiments of the present disclosure. [Figure 6] 1 is a flowchart of an example process for training one or more networks for a hybrid video codec, according to various embodiments of the present disclosure. [Figure 7] FIG. 1 illustrates an example of a learning-based video codec, according to various embodiments of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0012] The following detailed description of the exemplary embodiments refers to the accompanying drawings, in which the same reference numbers in different drawings may identify the same or similar elements.

[0013] The foregoing disclosure provides illustrations and descriptions, but is not intended to be exhaustive or to limit implementations to the precise disclosed implementations. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations. Further, one or more features or components of some embodiments may be combined or incorporated into some embodiments (or one or more features of some embodiments). Additionally, in the flowcharts and operational descriptions provided below, it is understood that one or more operations may be omitted, one or more operations may be added, one or more operations may be performed concurrently (at least in part), and the order of one or more operations may be switched.

[0014] It will be apparent that the systems and / or methods described herein may be implemented in different forms, such as hardware, firmware, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and / or methods is not intended to limit the implementation. Thus, the operation and behavior of the systems and / or methods are described herein without reference to specific software code, and it will be understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0015] Although particular feature combinations are recited in the claims and / or disclosed herein, these combinations are not intended to limit the disclosure of possible implementations. Indeed, many of these features may be combined in ways not specifically recited in the claims and / or disclosed herein. Although each dependent claim listed below may depend directly on only one claim, the disclosure of possible implementations includes each dependent claim in combination with all other claims in the claim set.

[0016] No element, act, or instruction used herein should be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles "a" and "an" are intended to include one or more items and may be used interchangeably with "one or more." When only one item is intended, the term "one" or similar language is used. Also, as used herein, terms such as "has," "have," "having," "include," and "including" are intended to be open-ended terms. Furthermore, the phrase "based on" is intended to mean "based at least in part on," unless otherwise specified. Furthermore, phrases such as "at least one of [A] and [B]" or "at least one of [A] or [B]" should be understood to include A only, B only, or both A and B.

[0017] Throughout this specification, references to "some embodiments," "one embodiment," or similar language mean that a particular feature, structure, or characteristic described in connection with the illustrated embodiment is included in some embodiments of the solution. Thus, throughout this specification, the phrases "some embodiments," "one embodiment," and similar language may, but do not necessarily, all refer to the same embodiments.

[0018] Furthermore, the described features, advantages, and characteristics of the present disclosure may be combined in any suitable manner in one or more embodiments. In light of the description herein, those skilled in the art will recognize that the present disclosure may be practiced without one or more of the specific features or advantages of a particular embodiment. In other instances, additional features and advantages may be recognized in certain embodiments that may not be present in all embodiments of the present disclosure.

[0019] The disclosed methods may be used separately or combined in any order. Furthermore, each of the methods (or embodiments), encoders, and decoders may be implemented by processing circuitry (e.g., one or more processors, or one or more integrated circuits). In one example, the one or more processors execute a program stored on a non-transitory computer-readable medium.

[0020]

[0001] Embodiments of the present disclosure relate to video coding for machines. Specifically, methods for encoding video for machine vision and hybrid human-machine vision are disclosed. Traditional video codecs are designed for human consumption. In some embodiments, traditional video codecs can be combined with learning-based codecs to form hybrid codecs, thereby efficiently encoding video for machine vision and hybrid human-machine vision.

[0021] 1 is a diagram of an exemplary device for performing a translation service. Device 100 may correspond to any type of known computer, server, or data processing device. For example, device 100 may comprise a processor, a personal computer (PC), a printed circuit board (PCB) with computing device, a minicomputer, a mainframe computer, a microcomputer, a telephone computing device, a wired / wireless computing device (e.g., a smartphone, a personal digital assistant (PDA)), a laptop, a tablet, a smart device, or any other similar operating device.

[0022] In some embodiments, as shown in FIG. 1, device 100 may include a set of components such as a processor 120, a memory 130, a storage component 140, an input component 150, an output component 160, and a communication interface 170.

[0023] Bus 110 may comprise one or more components that enable communication between a set of components of device 100. For example, bus 110 may be a communications bus, a crossover bar, a network, etc. Although bus 110 is shown in FIG. 1 as a single line, bus 110 may be implemented using multiple (two or more) connections between a set of components of device 100. The present disclosure is not limited in this respect.

[0024] Device 100 may include one or more processors, such as processor 120. Processor 120 may be implemented in hardware, firmware, and / or a combination of hardware and software. For example, processor 120 may include a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a general-purpose single-chip or multi-chip processor, or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the operations described herein. A general-purpose processor may be a microprocessor, or any conventional processor, controller, microcontroller, or state machine. Processor 120 may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. In some embodiments, particular processes and methods may be performed by circuitry specific to a given operation.

[0025] The processor 120 may control the overall operation of the device 100 and / or a set of components of the device 100 (e.g., memory 130, storage components 140, input components 150, output components 160, and communication interface 170).

[0026] Device 100 may further comprise memory 130. In some embodiments, memory 130 may include random access memory (RAM), read-only memory (ROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic memory, optical memory, and / or another type of dynamic or static storage device. Memory 130 may store information and / or instructions for use (e.g., execution) by processor 120.

[0027] Storage component 140 of device 100 may store information and / or computer-readable instructions and / or code related to the operation and use of device 100. For example, storage component 140 may include a hard disk (e.g., a magnetic disk, optical disk, magneto-optical disk, and / or solid-state disk), a compact disk (CD), a digital versatile disk (DVD), a universal serial bus (USB) flash drive, a Personal Computer Memory Card International Association (PCMCIA) card, a floppy disk, a cartridge, a magnetic tape, and / or another type of non-transitory computer-readable medium, along with a corresponding drive.

[0028] Device 100 may further comprise input component 150. Input component 150 may include one or more components that enable device 100 to receive information via user input (e.g., a touchscreen, a keyboard, a keypad, a mouse, a stylus, a button, a switch, a microphone, a camera, etc.), etc. Alternatively or additionally, input component 150 may include sensors for sensing information (e.g., a global positioning system (GPS) component, an accelerometer, a gyroscope, an actuator, etc.).

[0029] Output components 160 of device 100 may include one or more components that may provide output information from device 100 (e.g., a display, a liquid crystal display (LCD), a light emitting diode (LED), an organic light emitting diode (OLED), a haptic feedback device, a speaker, etc.).

[0030] Device 100 may further comprise a communication interface 170. Communication interface 170 may include a receiver component, a transmitter component, and / or a transceiver component. Communication interface 170 may enable device 100 to establish a connection and / or transfer communications with other devices (e.g., a server, another device). The communications may occur via a wired connection, a wireless connection, or a combination of wired and wireless connections. Communication interface 170 may enable device 100 to receive information from and / or provide information to another device. In some embodiments, communication interface 170 may provide communication with another device over a network such as a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a private network, an ad hoc network, an intranet, the Internet, an optical fiber-based network, a cellular network (e.g., a fifth generation (5G) network, a long term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a telephone network (e.g., a public switched telephone network (PSTN)), etc., and / or a combination of these or other types of networks. Alternatively or additionally, communication interface 170 may provide communication with another device over a device-to-device (D2D) communication link such as FlashLinQ, WiMedia, Bluetooth, ZigBee, Wi-Fi, LTE, 5G, etc. In other embodiments, communication interface 170 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, etc.

[0031] Device 100 may be included in core network 240 and may perform one or more processes described herein. Device 100 may perform operations based on processor 120 executing computer-readable instructions and / or code, which may be stored by a non-transitory computer-readable medium, such as memory 130 and / or storage component 140. A computer-readable medium may refer to a non-transitory memory device. A memory device may include memory space within a single physical storage device and / or memory space scattered across multiple physical storage devices.

[0032] Computer-readable instructions and / or code may be loaded into memory 130 and / or storage component 140 from another computer-readable medium or from another device via communications interface 170. The computer-readable instructions and / or code stored in memory 130 and / or storage component 140, when executed by processor 120, may cause device 100 to perform one or more processes described herein.

[0033] Alternatively, or in addition, hardwired circuitry may be used in place of or in combination with software instructions to implement one or more processes described herein. Thus, the embodiments described herein are not limited to any specific combination of hardware circuitry and software.

[0034] The number and arrangement of components shown in Figure 1 are provided as an example. In practice, there may be additional, fewer, different, or differently arranged components than those shown in Figure 1. Furthermore, two or more components shown in Figure 1 may be implemented within a single component, or a single component shown in Figure 1 may be implemented as multiple distributed components. Additionally or alternatively, a set of components shown in Figure 1 may perform one or more operations described as being performed by another set of components shown in Figure 1.

[0035] 2 is a block diagram illustrating one embodiment of a hybrid video codec 200. Codec200 may include a conventional codec 220 and a learning-based codec 230. The input 201 to the hybrid codec can be a video or an image, since an image can be treated as a special type of video (e.g., a video with one image). In FIG. 2, the conventional video codec 220 may be used to compress the input video 201 at different scales (e.g., the original resolution or downsampling), and the downsampling ratio of the downsampling module 210 may be fixed and known to both the encoder 221 and the decoder 223, or the downsampling ratio may be user-defined, e.g., 100% (e.g., no downsampling), 50%, 25%, etc., and may be transmitted as metadata in the bitstream 224 to inform the decoder 222. The conventional video codec may be an image codec such as VVC, HEVC, H264, JPEG, or JPEG 2000. The downsampling module 210 may be a classical image downsampler or a learning-based image downsampler. The decoded and downsampled video 203 (e.g., "low-resolution video 203" in FIG. 2) may be upsampled to the original resolution of the video (e.g., "high-resolution video 204") that can be used for human vision using an upsampling module 250. The upsampling module 250 may be a classical image upsampler or a learning-based image upsampler, such as a learning-based super-resolution module.

[0036] In some embodiments, the hybrid video codec 200 may also employ a learning-based video codec 230 to compress the downsampled video 202 .

[0037] The encoder may generate a reconstructed video 203 and upsample it to the original input resolution. The upsampled reconstructed video 205 may then be subtracted from the input video to generate a residual video signal 202, which may be fed to the learning-based codec 230 of FIG. 2. The upsampling module 240 of the hybrid video codec 200 may be the same as the upsampling module 250 (after the low resolution). The video 203 may be decoded at the decoder. The output of the residual decoder 238 may be added onto the high-resolution video 204 to form a reconstructed video 205, which may be used for machine vision tasks.

[0038] The networks of Figures 4 and 5, which are described in more detail below, can be trained using images. Disclosed networks, such as the network of Figure 6, need to be retrained using the residual signal 202 as shown in Figure 2.

[0039] During training, the input to the learning-based codec 230 may be the residual signal / image 202, which may also be the ground truth. The loss function may utilize a rate-distortion loss as follows: L overall =R+λ mse L mse (1)

[0040] In Equation 1, R represents the bitstream cost, which may be an estimated bits per pixel (BPP) value, and λ mse λ is the mean square error between the residual image 202 and the corresponding reconstructed residual image 205, as shown in FIG. mse is a positive weighting factor used to trade off bitrate cost against compression performance. In some embodiments, the rate-distortion loss may be modified as follows: L overall =R+λ ms-ssim L ms-ssim (2) where L ms-ssimis the MS-SSIM metric calculated using the residual image 202 and the corresponding reconstructed residual image 205, and λ ms-ssim is a weighting coefficient.

[0041] In some embodiments, a task network may be used in the joint training process. A machine vision task network, such as the object detection network YOLOv3 or Faster R-CNN, may be added for joint training. The high-resolution video 204 in FIG. 2 may be an upsampled and decoded video. detect is the detection loss calculated in the object detection network. The total loss function can be expressed as Equation 3 or 4. L overall =R+λ mse L mse +λ detect L detect (3) L overall =R+λ ms-ssim L ms-ssim +λ detect L detect (4) where λ detect is a positive weighting coefficient. During training, the model parameters for the machine vision network may be fixed, and only the parameters of the residual encoder 231 / decoder 238, entropy encoder 233 / decoder 236, and entropy model 235 are trained. In some embodiments, the parameters of the machine vision network may be trained together with the rest of the network of FIG. 2.

[0042] 3 illustrates an embodiment of an architecture for a video coding machine (VCM), such as hybrid video codec 200. Sensor output 300 travels a video coding path 311 through a VCM encoder 310 to a VCM decoder 320, where it undergoes video decoding 321. Alternative paths are for feature extraction 312, feature transformation 313, feature encoding 314, and feature decoding 322. The output of VCM decoder 320 is primarily intended for machine consumption, i.e., machine vision 305. In some cases, it can also be used for human vision 306. One or more machine tasks for understanding the video content are then performed.

[0043] 4 and 5 show the architecture of the learning-based image codec 230 in some embodiments. The learning-based codec 230 may be an image codec, so it may compress the residual video signal in the above section frame by frame without considering temporal redundancy between frames. For example, the learning-based image codec 230 may follow an autoencoder architecture as shown in FIGS. 4 and 5. In some examples, the inverse quantization modules 401A and 401B may be removed because the quantization operations 401A and 401B are rounding operations (e.g., rounding floating-point numbers to their nearest integers). The corresponding inverse quantization module 237 may be an identification module and may be removed from the architecture. Therefore, depending on the operation of the quantization module 232, the inverse quantization module 237 in FIG. 2 may be removed.

[0044] The two architectures in Figures 4 and 5 are similar autoencoder architectures. Figure 4 shows an example architecture that includes an analysis network 410, a synthesis network 420, quantization 401A, 401B, arithmetic coders 402A, 402B and decoders 403A, 403B, and an entropy model 430. The differences lie in the details of the analysis network 410, the synthesis network 420, and the entropy model 430.

[0045] In addition, the network may be designed for normal image compression. Therefore, the number of filters in each convolution module may be larger. For example, in FIG. 2, N=128 and M=192 or 320. Similarly, in FIG. 3, N=192. In FIG. 2, since the learning-based image codec 230 may be used to compress the residual signal 202, the number of filters in the convolution filter module may be significantly reduced to minimize performance degradation and reduce complexity. The GDN (Generalized Division and Normalization) module 511 and the IGDN (Inverse Generalized Division and Normalization) module 521 may be replaced with the ReLU (Registered Linear Unit) 531 in FIG. 5. To reduce complexity, the attention module in FIG. 4 may be removed, and the Gaussian Mixture Entropy Model may be replaced with the scaled hyper-priori module shown in FIG. 2.

[0046] In this disclosure, the architectures of Figures 4 and 5 are used as examples, and any autoencoder that follows the spirit of Figures 4 or 5, or reduced-complexity versions thereof, may be utilized. For example, an example of such an architecture is shown in Figure 2. In some embodiments, the learning-based codec 230 may be a video codec, such that inter-frame temporal redundancy in the residual signal 202 may be utilized. An example of a learning-based video codec 230 is shown in Figure 7.

[0047] FIG. 6 shows a flowchart disclosing one embodiment of a process for training one or more networks. The process may begin at operation S610, where an input including at least one video or image is received at a hybrid codec. The hybrid codec may include a first and a second codec, where the first codec is a conventional codec designed for human consumption and the second codec is a learning-based codec designed for machine vision. As an example, the hybrid codec may be hybrid video codec 200, which includes a first and a second codec. The first codec may be conventional codec 220, and the second codec may be learning-based codec 230, which receives input 201 as seen in FIG. 2. The process continues at operation S620, where the input is compressed using the first codec. For example, input 201 may be a first Codec 220. The compressing may include downsampling the input using downsampling module 210 and upsampling the compressed input using upsampling module 240, which generates residual signal 202. The process proceeds to operation S630, where the residual signal (e.g., residual signal 202) is quantized to obtain a quantized representation of the input using a quantizer, such as quantizer 232. The process proceeds to operation S640, where the quantized representation of the input is entropy coded using one or more convolution filter modules of an entropy model, such as entropy model 235. The process proceeds to operation S650, where one or more networks are trained using the entropy coded quantized representation of the input and the reconstructed video, such as reconstructed video 205.

[0048] To specify the hybrid video codec 200, several parameters need to be specified through a high-level syntax, such as a sequence parameter set, a picture parameter set, a picture header, etc. Alternatively, such information can be delivered through the system level or via metadata using SEI messages.

[0049] Downsampling ratio: A set of downsampling ratios, e.g., {r0,r1,…,r N-1} may be defined, where N is the number downsampling ratio. Let W and H be the height and width of the input image resolution, and let r be the downsampling ratio. n Then the height and width of the downsampled image are r n W and r n H. The index n∈{0,1,…N-1} is the sampling rate r used in the hybrid video codec. p The decoder uses the upsampling ratio to upsample the decoded low-resolution image / video to obtain a high-resolution image / video.

number

[0050] Upsampling Module (Upsampler): If only one type of upsampler is used in a hybrid video codec, there is no need to specify the upsampler in the bitstream. However, if multiple types of upsamplers, each with different complexity and performance, can be employed in the codec, information about the type of upsampler needs to be specified in the bitstream. For example, if there are M types of upsamplers, an index m∈{0, 1, ... M-1} is used to specify which type of upsampler is used. The index m can be binarized using a fixed-length code or a p-th order Exponential-Golomb code and transmitted in the bitstream using bypass coding. In some embodiments, p=0 or 1.

[0051] Codec for encoding downsampled images / videos: If only one type of codec is used in a hybrid video code for encoding downsampled images / videos, there is no need to specify the codec in the bitstream. However, if multiple types of codecs, such as VVC, HEVC, and H264, can be adopted in the system, information about the type of codec needs to be specified in the bitstream to enable proper decoding at the decoder. For example, if there are Q types of upsamplers, an index q∈{0, 1, ... Q-1} is used to specify which type of codec is used. The index q can be binarized using a fixed-length code or a p-th order exponential-Golomb code and transmitted in the bitstream using bypass coding. In some embodiments, p=0 or 1.

[0052] Codec for Encoding Residual Image / Video: Similarly, if only one type of learning-based codec is used in a hybrid video codec for encoding residual image / video, there is no need to specify the codec in the bitstream. However, if multiple types of codecs can be employed in the system, such as the codecs shown in Figures 2 to 4, information about the type of codec needs to be specified in the bitstream to enable proper decoding at the decoder. For example, if there are L types of codecs, an index l ∈ {0, 1, ... L-1} can be used to specify which type of codec is used. The index l can be binarized using a fixed-length code or a p-th order exponential-Golomb code and transmitted in the bitstream using bypass coding. In some embodiments, p = 0 or 1. The different types of codecs should be known by both the encoder and the decoder.

[0053] In some embodiments, instead of using an index, a description of the network structure of the learning-based codec can be specified in a high-level syntax or metadata. For example, individual modules in the decoder network can be specified, such as a 3x3 convolution module with N output filters, a 3x3 convolution module with N output filters with 2x upsampling, etc. In addition to the network structure, corresponding decoder model parameters can be transmitted in the bitstream to enable the decoder to decode the bitstream and generate a reconstructed residual image / video. The decoder network must be symmetric with the encoder network. Partial reconstruction of the residual image / video is possible. In addition, a fixed-point implementation can be specified instead of a floating-point implementation of the learning-based codec.

[0054] Parameter Set Selection for Different Machine Tasks: In some embodiments, the residual video signal coding branch may train different parameter sets for different machine vision tasks while the same network architecture is shared by the different tasks. If this framework supports more than one task to perform residual coding for one specific video input, a target machine task should be specified. For example, if there are T types of machine vision tasks, an index t ∈ {0, 1, ... T-1} may be used to specify which type of machine task to target. Thus, the residual signal coding branch switches to the corresponding parameter set trained for that task.

[0055] Quantization: Quantization generally refers to dividing continuously varying data into a finite number of division levels and assigning a specific value to each level. The most basic form of quantization is uniform quantization, which has quantization intervals of the same size within a certain range. For example, one method is to set the size of the quantization interval by dividing the minimum and maximum values ​​for a particular input data by the desired quantization bits.

[0056] The foregoing disclosure provides illustration and description, but is not intended to be exhaustive or to limit implementations to the precise implementations disclosed. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations.

[0057] It is understood that the specific order or hierarchy of blocks in the processes / flowcharts disclosed herein is illustrative of example approaches. Based on design preferences, it is understood that the specific order or hierarchy of blocks in the processes / flowcharts may be rearranged. Also, some blocks may be combined or omitted. The accompanying method claims present elements of the various blocks in a sample order and are not meant to be limited to the specific order or hierarchy presented.

[0058] Some embodiments may relate to a system, a method, and / or a computer-readable medium at any possible level of integration of technical details. Furthermore, one or more of the above components may be implemented as instructions stored on a computer-readable medium and executable by at least one processor (and / or may include at least one processor). The computer-readable medium may include a computer-readable non-transitory storage medium having computer-readable program instructions for causing the processor to perform operations.

[0059] A computer-readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. The computer-readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. A non-exhaustive list of more specific examples of computer-readable storage media includes the following: portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically encoded devices such as punch cards or ridge-in-groove structures with instructions recorded thereon, and any suitable combination of the above. As used herein, computer-readable storage media should not be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses passing through a fiber optic cable), or electrical signals transmitted over electrical wires.

[0060] The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to each computing / processing device, or may be downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium within the respective computing / processing device.

[0061] The computer-readable program code / instructions for carrying out operations may be either assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data for integrated circuits, or source or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk or C++, and procedural programming languages ​​such as the "C" programming language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet Service Provider). In some embodiments, an electronic circuit, including, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may execute computer-readable program instructions by utilizing state information of the computer-readable program instructions to personalize the electronic circuit to perform an aspect or operation.

[0062] These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create means for implementing the operations specified in the flowchart and / or block diagram blocks. These computer-readable program instructions can also be stored on a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and / or other device to operate in a particular manner, such that the computer-readable storage medium having stored instructions comprises an article of manufacture containing instructions that implement aspects of the operations specified in the flowchart and / or block diagram blocks.

[0063] The computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device such that a sequence of operations are executed on the computer, other programmable apparatus, or other device to generate a computer-implemented process, such that the instructions, executing on the computer, other programmable apparatus, or other device, implement the operations specified in the flowchart and / or block diagram blocks.

[0064] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer-readable media according to various embodiments. In this regard, each block in the flowcharts or block diagrams may represent a module, segment, or portion of instructions, including one or more executable instructions for implementing the specified logical operations. The methods, computer systems, and computer-readable media may include additional, fewer, different, or differently arranged blocks than those shown in the figures. In some alternative implementations, the operations in the blocks may be performed in a different order than shown in the figures. For example, two blocks shown in succession may actually be executed concurrently or substantially concurrently, or may even be executed in the reverse order, depending on the functionality involved. Each block in the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams, may be implemented by a special-purpose hardware-based system that performs the specified operations or executes a combination of special-purpose hardware and computer instructions.

[0065] It will be apparent that the systems and / or methods described herein may be implemented in different forms, such as hardware, firmware, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and / or methods is not intended to limit the implementation. Thus, the operation and behavior of the systems and / or methods are described herein without reference to specific software code, and it will be understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0066] The above disclosure also encompasses the embodiments listed below.

[0067] (1) A method for encoding video for machine vision and human / machine hybrid vision, the method being executed by one or more processors and including: receiving an input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision; compressing the input using the first codec, the compressing including downsampling the input using a downsampling module and upsampling the compressed input using an upsampling module to generate a residual signal; quantizing the residual signal to obtain a quantized representation of the input; entropy encoding the quantized representation of the input using one or more convolutional filter modules; and training one or more networks using the entropy-encoded quantized representation.

[0068] (2) The method according to feature (1), wherein the conventional codec includes any of the VVC, HEVC, H264, JPEG, or JPEG2000 codecs.

[0069] (3) The method of features (1) or (2), wherein the training codec includes an image codec, and the residual signal is compressed frame by frame without considering temporal redundancy.

[0070] (4) The method of any one of features (1) to (3), wherein the downsampling module is one of a classical image downsampler and a learning-based image downsampler.

[0071] (5) The method of any one of features (1) to (4), wherein the downsampling module uses a downsampling ratio N that is fixed and known at both the encoder and decoder, or that is user-defined.

[0072] (6) The method of any one of features (1) to (5), wherein the upsampling module is one of a classical image upsampler and a learning-based image upsampler.

[0073] (7) The method of any one of features (1) to (6), wherein the upsampled compressed input is subtracted from the input to generate a second residual signal, and the second residual signal is provided to the learning-based codec.

[0074] (8) The method of feature (7), wherein the output of the second codec is added on top of the upsampled compressed input to form a reconstructed video suitable for machine vision tasks.

[0075] (9) The method of any one of features (1) to (8), wherein the input to the hybrid codec is ground truth.

[0076] (10) The method of any one of features (1) to (9), wherein the machine vision network is fixed and parameters of the residual coder, entropy coder, and entropy model of the second codec are trained.

[0077] (11) An apparatus for encoding video for machine vision and human / machine hybrid vision, comprising: at least one memory configured to store computer program code; and at least one processor configured to access the computer program code and operate as instructed by the computer program code, the computer program code causing the at least one processor to receive an input, including at least one of video or image data, at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision; and configuration code configured to cause the at least one processor to compress the input using the first codec. and at least one processor including: a compression code configured to: downsample an input using a downsampling module; and an upsampling code configured to cause the at least one processor to upsample the compressed input using an upsampling module to generate a residual signal; a quantization code configured to cause the at least one processor to quantize the residual signal to obtain a quantized representation of the input; an entropy coding code configured to cause the at least one processor to encode the quantized representation of the input using one or more convolutional filter modules; and a training code configured to train one or more networks using the entropy coded quantized representations.

[0078] (12) The apparatus of feature (11), wherein the conventional codec includes at least one of a VVC codec, an HEVC codec, an H264 codec, a JPEG codec, or a JPEG2000 codec.

[0079] (13) The apparatus of feature (11) or (12), wherein the training codec includes an image codec, and the residual signal is compressed frame by frame without considering temporal redundancy.

[0080] (14) The apparatus of claim 11, wherein the downsampling module is one of a classical image downsampler and a learning-based image downsampler. (11) The apparatus of any one of (11) to (13).

[0081] (15) The apparatus of any one of features (11) to (14), wherein the downsampling module uses a downsampling ratio N that is fixed and known at both the encoder and decoder, or that is user-defined.

[0082] (16) The apparatus according to any one of features (11) to (15), wherein the upsampling module is one of a classical image upsampler and a learning-based image upsampler.

[0083] (17) The apparatus of any one of features (11) to (16), wherein the upsampled compressed input is subtracted from the input to generate a second residual signal, and the second residual signal is provided to the learning-based codec.

[0084] (18) The apparatus of feature (17), wherein the output of the second codec is added on top of the upsampled compressed input to form a reconstructed video configured for a machine vision task.

[0085] (19) The apparatus of any one of features (11) to (18), wherein the input to the hybrid codec is ground truth.

[0086] (20) A non-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, cause the processor to perform a method for encoding video for machine vision and human / machine hybrid vision, the method including: receiving an input including at least one of video or image data at a hybrid codec, the hybrid codec including first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision; compressing the input using the first codec, the compressing including downsampling the input using a downsampling module and upsampling the compressed input using an upsampling module to generate a residual signal; quantizing the residual signal to obtain a quantized representation of the input; entropy encoding the quantized representation of the input using one or more convolutional filter modules; and training one or more networks using the entropy-encoded quantized representation. [Explanation of symbols]

[0087] 100 devices 110 Bus 120 processors 130 memory 140 Memory Components 150 Input Components 160 Output Components 170 Communication Interface 200 Hybrid Video Codec 201 Input Video 202 Downsampled Video 203 Decoded and downsampled video, low resolution video 204 High Definition Video 205 Reconstructed Video 210 Downsampling Module 220 traditional video codec 221 Encoder 222 decoder 223 decoder 224 bitstream 230 Learning-Based Video or Image Codec 231 Residual Encoder 232 Quantizer, Quantization Module 233 Entropy Encoder 235 Entropy Model 236 Entropy Decoder 237 Inverse Quantization Module 238 Residual Decoder 240 Core network, upsampling module 250 Upsampling Module 300 sensor output 305 Machine Vision 306 Human Vision 310 Video Coding (VCM) Encoder for Machines 311 Video Encoding Path 312 Feature Extraction 313 Feature Transformation 314 Feature Coding 320 VCM decoder 321 video decoding 322 Feature Decoding 401A Quantization 401B Quantization 402A Arithmetic Coda 402B Arithmetic Coda 403A Arithmetic Decoder 403B Arithmetic Decoder 410 Analysis Network 420 Synthetic Network 430 Entropy Model 511 Generalized Division Normalization (GDN) Module 521 Inverse Generalized Division Normalization (IGDN) Module

Claims

1. 1. A method for encoding video for machine vision and human / machine hybrid vision, the method being executed by one or more processors and comprising: The method comprises: receiving an input comprising at least one of video or image data at a hybrid codec, the hybrid codec comprising first and second codecs, the first codec being a conventional codec designed for human consumption and the second codec being a learning-based codec designed for machine vision; compressing the input using the first codec, the compressing step comprising downsampling the input using a downsampling module; upsampling the compressed input using an upsampling module that generates a residual signal; quantizing the residual signal to obtain a quantized representation of the input; entropy encoding the quantized representation of the input using one or more convolution filter modules; training one or more networks using the entropy coded quantized representations; Including, the upsampled compressed input is subtracted from the input to generate the residual signal; the residual signal is provided to the learning-based codec; the output of the second codec is added on top of the upsampled compressed input to form a reconstructed video suitable for machine vision tasks; The method, wherein the step of training one or more networks using the entropy-encoded quantized representations includes determining a value of an index that specifies what type of the machine vision task is a target of the training step.

2. The method of claim 1 , wherein the legacy codec comprises any of a VVC, HEVC, H264, JPEG, or JPEG2000 codec.

3. The method of claim 1 , wherein the learning-based codec comprises an image codec, and the residual signal is compressed frame-by-frame without considering temporal redundancy.

4. The method of claim 1 , wherein the downsampling module is one of a classical image downsampler and a learning-based image downsampler.

5. The method of claim 4 , wherein the downsampling module uses a downsampling ratio N that is fixed and known at both the encoder and the decoder, or that is user-defined.

6. The method of claim 1 , wherein the upsampling module is one of a classical image upsampler and a learning-based image upsampler.

7. The method of claim 1 , wherein the input to the hybrid codec is ground truth.

8. The method of claim 1 , wherein a machine vision network is fixed and parameters of a residual coder, an entropy coder, and an entropy model of the second codec are trained.

9. Apparatus configured to perform the method according to any one of claims 1 to 8.

10. A computer program product for causing at least one processor to carry out the method according to any one of claims 1 to 8.

Images