Method, apparatus, and medium for video processing

Abstract

Embodiments of the present disclosure provide a solution for video processing. A method for video processing is proposed. The method comprises: determining, for a conversion between a current video unit of a video and a bitstream of the video with a neural network (NN)-based model comprising a first module and a second module, a first latent representation associated with motion information of the current video unit, by the first module with a first plurality of parameters; determining, by the second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and performing the conversion based on the first latent representation and the second latent representation, wherein the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

Description

METHOD, APPARATUS, AND MEDIUM FOR VIDEO PROCESSING FIELDS

[0001] Embodiments of the present disclosure relates generally to video processing techniques, and more particularly, to neural network-based video coding. BACKGROUND

[0002] The past decade has witnessed the rapid development of deep learning in a variety of areas, especially in computer vision and image processing. Neural network was invented originally with the interdisciplinary research of neuroscience and mathematics. It has shown strong capabilities in the context of non-linear transform and classification. Neural network- based image / video compression technology has gained significant progress during the past half decade. It is reported that the latest neural network-based image compression algorithm achieves comparable rate-distortion (R-D) performance with Versatile Video Coding (VVC). With the performance of neural image compression continually being improved, neural network-based video compression has become an actively developing research area. However, coding performance of neural network-based image / video coding is generally expected to be further improved. SUMMARY

[0003] Embodiments of the present disclosure provide a solution for video processing.

[0004] In a first aspect, a method for video processing is proposed. The method comprises: determining, for a conversion between a current video unit of a video and a bitstream of the video with a neural network (NN)-based model comprising a first module and a second module, a first latent representation associated with motion information of the current video unit, by the first module with a first plurality of parameters; determining, by the second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and performing the conversion based on the first latent representation and the second latent representation, wherein the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0005] Based on the method in accordance with the first aspect of the present disclosure, the first module for generating the first latent representation associated with motion information is independent from the second module for generating the second latent representation associated with residual information of the current video unit. Compared with the conventional solution, the proposed method can advantageously decouple the first latent 1 F1254298PCTrepresentation and the second latent representation. Thereby, the proposed method can advantageously improve the coding performance.

[0006] In a second aspect, an apparatus for video processing is proposed. The apparatus comprises a processor and a non-transitory memory with instructions thereon. The instructions upon execution by the processor, cause the processor to perform a method in accordance with the first aspect of the present disclosure.

[0007] In a third aspect, a non-transitory computer-readable storage medium is proposed. The non-transitory computer-readable storage medium stores instructions that cause a processor to perform a method in accordance with the first aspect of the present disclosure.

[0008] In a fourth aspect, another non-transitory computer-readable recording medium is proposed. The non-transitory computer-readable recording medium stores a bitstream of a video which is generated by a method performed by an apparatus for video processing. The method comprises: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and generating, by a neural network (NN)-based model, the bitstream based on the first latent representation and the second latent representation, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0009] In a fifth aspect, a method for storing a bitstream of a video is proposed. The method comprises: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; generating, by a neural network (NN)-based model, the bitstream based on the first latent representation and the second latent representation; and storing the bitstream in a non-transitory computer- readable recording medium, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0010] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended 2 F1254298PCTto identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Through the following detailed description with reference to the accompanying drawings, the above and other objectives, features, and advantages of example embodiments of the present disclosure will become more apparent. In the example embodiments of the present disclosure, the same reference numerals usually refer to the same components.

[0012] Fig. 1 illustrates a block diagram that illustrates an example visual data coding system, in accordance with some embodiments of the present disclosure;

[0013] Fig. 2 is a schematic diagram illustrating an example transform coding scheme;

[0014] Fig.3 illustrates an inter-picture decoding process of cool-chic video for low-delay application;

[0015] Fig. 4 illustrates an inter-picture decoding process in accordance with some embodiments of the present disclosure;

[0016] Fig. 5 illustrates a flowchart of a method for video processing in accordance with some embodiments of the present disclosure; and

[0017] Fig. 6 illustrates a block diagram of a computing device in which various embodiments of the present disclosure can be implemented.

[0018] Throughout the drawings, the same or similar reference numerals usually refer to the same or similar elements. DETAILED DESCRIPTION

[0019] Principle of the present disclosure will now be described with reference to some embodiments. It is to be understood that these embodiments are described only for the purpose of illustration and help those skilled in the art to understand and implement the present disclosure, without suggesting any limitation as to the scope of the disclosure. The disclosure described herein can be implemented in various manners other than the ones described below.

[0020] In the following description and claims, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure belongs.

[0021] References in the present disclosure to “one embodiment,” “an embodiment,” “an example embodiment,” and the like indicate that the embodiment described may include a particular feature, structure, or characteristic, but it is not necessary that every embodiment includes the particular feature, structure, or characteristic. Moreover, such phrases are not 3 F1254298PCTnecessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an example embodiment, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.

[0022] It shall be understood that although the terms “first” and “second” etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and similarly, a second element could be termed a first element, without departing from the scope of example embodiments. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises”, “comprising”, “has”, “having”, “includes” and / or “including”, when used herein, specify the presence of stated features, elements, and / or components etc., but do not preclude the presence or addition of one or more other features, elements, components and / or combinations thereof. Example Environment

[0024] Fig. 1A is a block diagram that illustrates an example visual data coding system 100 that may utilize the techniques of this disclosure. As shown, the visual data coding system 100 may include a source device 110 and a destination device 120. The source device 110 can be also referred to as a visual data encoding device, and the destination device 120 can be also referred to as a visual data decoding device. In operation, the source device 110 can be configured to generate encoded visual data and the destination device 120 can be configured to decode the encoded visual data generated by the source device 110. The source device 110 may include a visual data source 112, a visual data encoder 114, and an input / output (I / O) interface 116.

[0025] The visual data source 112 may include a source such as a visual data capture device. Examples of the visual data capture device include, but are not limited to, an interface to receive visual data from a visual data provider, a computer graphics system for generating visual data, and / or a combination thereof.

[0026] The visual data may comprise one or more pictures of a video or one or more images. The visual data encoder 114 encodes the visual data from the visual data source 112 to 4 F1254298PCTgenerate a bitstream. The bitstream may include a sequence of bits that form a coded representation of the visual data. The bitstream may include coded pictures and associated visual data. The coded picture is a coded representation of a picture. The associated visual data may include sequence parameter sets, picture parameter sets, and other syntax structures. The I / O interface 116 may include a modulator / demodulator and / or a transmitter. The encoded visual data may be transmitted directly to destination device 120 via the I / O interface 116 through the network 130A. The encoded visual data may also be stored onto a storage medium / server 130B for access by destination device 120.

[0027] The destination device 120 may include an I / O interface 126, a visual data decoder 124, and a display device 122. The I / O interface 126 may include a receiver and / or a modem. The I / O interface 126 may acquire encoded visual data from the source device 110 or the storage medium / server 130B. The visual data decoder 124 may decode the encoded visual data. The display device 122 may display the decoded visual data to a user. The display device 122 may be integrated with the destination device 120, or may be external to the destination device 120 which is configured to interface with an external display device.

[0028] The visual data encoder 114 and the visual data decoder 124 may operate according to a visual data coding standard, such as video coding standard or still picture coding standard and other current and / or further standards.

[0029] Some exemplary embodiments of the present disclosure will be described in detailed hereinafter. It should be understood that section headings are used in the present document to facilitate ease of understanding and do not limit the embodiments disclosed in a section to only that section. Furthermore, while certain embodiments are described with reference to Versatile Video Coding or other specific visual data codecs, the disclosed techniques are applicable to other coding technologies also. Furthermore, while some embodiments describe coding steps in detail, it will be understood that corresponding steps decoding that undo the coding will be implemented by a decoder. Furthermore, the term visual data processing encompasses visual data coding or compression, visual data decoding or decompression and visual data transcoding in which visual data are represented from one compressed format into another compressed format or at a different compressed bitrate. 1. Brief Summary The present application relates to video compression using neural networks, and more particularly, to neural compression from a single video. It may be applied for low-latency or random-access scenarios. It may also be applicable to hybrid compression framework, where neural network- based coding tools are integrated into the framework of the existing video coding standards (e.g., 5 F1254298PCTHEVC, VVC) or future video coding standards. The described techniques relate to neural codecs that overfit a single instance (e.g. picture, video, etc.). The proposed techniques enhance the exploitation of temporal context in existing single instance-based neural compression by incorporating a well-designed, complexity-constrained motion component. Simple and effective training strategies are also developed to facilitate the learning of this motion component. 2. Introduction The past decade has witnessed the rapid development of deep learning in a variety of areas, especially in computer vision and image processing. Inspired from the great success of deep learning technology to computer vision areas, many researchers have shifted their attention from conventional image / video compression techniques to neural image / video compression technologies. Neural network was invented originally with the interdisciplinary research of neuroscience and mathematics. It has shown strong capabilities in the context of non-linear transform and classification. Neural network-based image / video compression technology has gained significant progress during the past half decade. It is reported that the latest neural network- based image compression algorithm achieves comparable R-D performance with Versatile Video Coding (VVC), the latest video coding standard developed by Joint Video Experts Team (JVET) with experts from MPEG and VCEG. While recent neural compression methods demonstrate strong rate-distortion (RD) performance, their high decoding complexity limits their applicability in scenarios with stringent constraints, particularly on mobile or edge devices with limited processing power. An alternative approach to reducing decoding complexity involves overfitting the neural codec to a single instance . In this method, the neural model is tailored to encode only a single video, which generally exhibits lower variance compared to a diverse set of natural videos. 2.1 Image / Video compression Image / video compression usually refers to the computing technology that compresses image / video into binary code to facilitate storage and transmission. The binary codes may or may not support losslessly reconstructing the original image / video, termed lossless compression and lossy compression. Most of the efforts are devoted to lossy compression since lossless reconstruction is not necessary in most scenarios. Usually the performance of image / video compression algorithms is evaluated from two aspects, i.e. compression ratio and reconstruction quality. Compression ratio is directly related to the number of binary codes, the less the better; Reconstruction quality is measured by comparing the reconstructed image / video with the original image / video, the higher the better. 6 F1254298PCTImage / video compression techniques can be divided into two branches, the classical video coding methods and the neural-network-based video compression methods. Classical video coding schemes adopt transform-based solutions, in which researchers have exploited statistical dependency in the latent variables (e.g., DCT or wavelet coefficients) by carefully hand- engineering entropy codes modeling the dependencies in the quantized regime. Neural network- based video compression is in two flavors, neural network-based coding tools and end-to-end neural network-based video compression. The former is embedded into existing classical video codecs as coding tools and only serves as part of the framework, while the latter is a separate framework developed based on neural networks without depending on classical video codecs. In the last three decades, a series of classical video coding standards have been developed to accommodate the increasing visual content. The international standardization organizations ISO / IEC has two expert groups namely Joint Photographic Experts Group (JPEG) and Moving Picture Experts Group (MPEG), and ITU-T also has its own Video Coding Experts Group (VCEG) which is for standardization of image / video coding technology. The influential video coding standards published by these organizations include JPEG, JPEG 2000, H.262, H.264 / AVC and H.265 / HEVC. After H.265 / HEVC, the Joint Video Experts Team (JVET) formed by MPEG and VCEG has been working on a new video coding standard Versatile Video Coding (VVC). The first version of VVC was released in July 2020. An average of 50% bitrate reduction is reported by VVC under the same visual quality compared with HEVC. Neural network-based image / video compression is not a new invention since there were a number of researchers working on neural network-based image coding. But the network architectures were relatively shallow, and the performance was not satisfactory. Benefit from the abundance of data and the support of powerful computing resources, neural network-based methods are better exploited in a variety of applications. At present, neural network-based image / video compression has shown promising improvements, confirmed its feasibility. Nevertheless, this technology is still far from mature and a lot of challenges need to be addressed. 2.2 Neural networks Neural networks, also known as artificial neural networks (ANN), are the computational models used in machine learning technology which are usually composed of multiple processing layers and each layer is composed of multiple simple but non-linear basic computational units. One benefit of such deep networks is believed to be the capacity for processing data with multiple levels of abstraction and converting data into different kinds of representations. Note that these representations are not manually designed; instead, the deep network including the processing layers is learned from massive data using a general machine learning procedure. Deep learning 7 F1254298PCTeliminates the necessity of handcrafted representations, and thus is regarded useful especially for processing natively unstructured data, such as acoustic and visual signal, whilst processing such data has been a longstanding difficulty in the artificial intelligence field. 2.3 Neural networks for image compression Existing neural networks for image compression methods can be classified in two categories, i.e., pixel probability modeling and auto-encoder. The former one belongs to the predictive coding strategy, while the latter one is the transform-based solution. Sometimes, these two methods are combined together in literature. 2.3.1 Pixel Probability Modeling According to Shannon’s information theory, the optimal method for lossless coding can reach theminimal coding rate − logଶ ^^(^^) where ^^(^^) is the probability of symbol ^^. A number of losslesscoding methods were developed in literature and among them arithmetic coding is believed to be among the optimal ones. Given a probability distribution ^^(^^), arithmetic coding ensures that thecoding rate to be as close as possible to its theoretical limit − logଶ ^^(^^) without considering therounding error. Therefore, the remaining problem is to how to determine the probability, which is however very challenging for natural image / video due to the curse of dimensionality. Following the predictive coding strategy, one way to model ^^(^^) is to predict pixel probabilitiesone by one in a raster scan order based on previous observations, where ^^ is an image.^^(^^) = ^^(^^^)^^(^^ଶ|^^^) …^^(^^^|^^^, … , ^^^ି^) … ^^(^^^×^|^^^, … , ^^^×^ି^) (1)where ^^ and ^^ are the height and width of the image, respectively. The previous observation is also known as the context of the current pixel. When the image is large, it can be difficult toestimate the conditional probability, thereby a simplified method is to limit the range of its context.^^(^^) = ^^(^^^)^^(^^ଶ|^^^) …^^(^^^|^^^ି^, … , ^^^ି^) … ^^(^^^×^|^^^×^ି^, … , ^^^×^ି^) (2)where ^^ is a pre-defined constant controlling the range of the context. It should be noted that the condition may also take the sample values of other color components into consideration. For example, when coding the RGB color component, R sample is dependent on previously coded pixels (including R / G / B samples), the current G sample may be coded according to previously coded pixels and the current R sample, while for coding the current B sample, the previously coded pixels and the current R and G samples may also be taken into consideration. Neural networks were originally introduced for computer vision tasks and have been proven to be effective in regression and classification problems. Therefore, it has been proposed using neuralnetworks to estimate the probability of ^^(^^^) given its context ^^^, ^^ଶ, … , ^^^ି^. In an existing design,the pixel probability is proposed for binary images, i.e., ^^^ ∈ {−1, +1}. The neural autoregressive8 F1254298PCTdistribution estimator (NADE) is designed for pixel probability modeling, where is a feed-forward network with a single hidden layer. A similar work is presented in an existing design, where the feed-forward network also has connections skipping the hidden layer, and the parameters are also shared. Experiments are performed on the binarized MNIST dataset. In an existing design, NADEis extended to a real-valued model RNADE, where the probability ^^(^^^|^^^, … , ^^^ି^) is derivedwith a mixture of Gaussians. Their feed-forward network also has a single hidden layer, but the hidden layer is with rescaling to avoid saturation and uses rectified linear unit (ReLU) instead of sigmoid. In an existing design, NADE and RNADE are improved by using reorganizing the order of the pixels and with deeper neural networks. Designing advanced neural networks plays an important role in improving pixel probability modeling. In an existing design, multi-dimensional long short-term memory (LSTM) is proposed, which is working together with mixtures of conditional Gaussian scale mixtures for probability modeling. LSTM is a special kind of recurrent neural networks (RNNs) and is proven to be good at modeling sequential data. The spatial variant of LSTM is used for images later in an existing design. Several different neural networks are studied, including RNNs and CNNs namely PixelRNN and PixelCNN, respectively. In PixelRNN, two variants of LSTM, called row LSTM and diagonal BiLSTM are proposed, where the latter is specifically designed for images. PixelRNN incorporates residual connections to help train deep neural networks with up to 12 layers. In PixelCNN, masked convolutions are used to suit for the shape of the context. Comparing with previous works, PixelRNN and PixelCNN are more dedicated to natural images: they consider pixels as discrete values (e.g., 0, 1, …, 255) and predict a multinomial distribution over the discrete values; they deal with color images in RGB color space; they work well on large-scale image dataset ImageNet. In an existing design, Gated PixelCNN is proposed to improve the PixelCNN, and achieves comparable performance with PixelRNN but with much less complexity. In an existing design, PixelCNN++ is proposed with the following improvements upon PixelCNN: a discretized logistic mixture likelihood is used rather than a 256-way multinomial distribution; down-sampling is used to capture structures at multiple resolutions; additional short-cut connections are introduced to speed up training; dropout is adopted for regularization; RGB is combined for one pixel. In an existing design, PixelSNAIL is proposed, in which casual convolutions are combined with self-attention. Most of the above methods directly model the probability distribution in the pixel domain. Some researchers also attempt to model the probability distribution as a conditional one upon explicit orlatent representations. That being said, it may be estimated that9 F1254298PCTwhere ^^ is the additional condition and ^^(^^) = ^^(^^)^^(^^|^^), meaning the modeling is split intoan unconditional one and a conditional one. The additional condition can be image label information or high-level representations. 2.3.2 Auto-encoder Auto-encoder originates from the well-known work proposed. The method is trained for dimensionality reduction and consists of two parts: encoding and decoding. The encoding part converts the high-dimension input signal to low-dimension representations, typically with reduced spatial size but a greater number of channels. The decoding part attempts to recover the high- dimension input from the low-dimension representation. Auto-encoder enables automated learning of representations and eliminates the need of hand-crafted features, which is also believed to be one of the most important advantages of neural networks. Fig. 2 illustrates a typical transform coding scheme. The original image x is transformed by the analysis network gୟto achieve the latent representation y. The latent representation y is quantized and compressed into bits. The number of bits R is used to measure the coding rate. The quantizedlatent representation y^ is then inversely transformed by a synthesis network g^ to obtain thereconstructed image x^. The distortion is calculated in a perceptual space by transforming x and x^ with the function g୮. It is intuitive to apply auto-encoder network to lossy image compression. it only need to encode the learned latent representation from the well-trained neural networks. However, it is not trivial to adapt auto-encoder to image compression since the original auto-encoder is not optimized for compression thereby not efficient by directly using a trained auto-encoder. In addition, there exist other major challenges: First, the low-dimension representation should be quantized before being encoded, but the quantization is not differentiable, which is required in backpropagation while training the neural networks. Second, the objective under compression scenario is different since both the distortion and the rate need to be take into consideration. Estimating the rate is challenging. Third, a practical image coding scheme needs to support variable rate, scalability, encoding / decoding speed, interoperability. In response to these challenges, a number of researchers have been actively contributing to this area. The prototype auto-encoder for image compression is in Fig. 2, which can be regarded as atransform coding strategy. The original image ^^ is transformed with the analysis network ^^ =^^^(^^), where ^^ is the latent representation which will be quantized and coded. The synthesisnetwork will inversely transform the quantized latent representation ^ ^^ back to obtain thereconstructed image ^ ^^ = ^^^(^^^). The framework is trained with the rate-distortion loss function,i.e., ℒ = ^^ + ^^^^, where ^^ is the distortion between ^^ and ^ ^^, ^^ is the rate calculated or estimated10 F1254298PCTfrom the quantized representation^ ^^, and ^^ is the Lagrange multiplier. It should be noted that ^^ can be calculated in either pixel domain or perceptual domain. All existing research works follow this prototype and the difference might only be the network structure or loss function. In terms of network structure, RNNs and CNNs are the most widely used architectures. In the RNNs relevant category, Toderici et al. propose a general framework for variable rate image compression using RNN. They use binary quantization to generate codes and do not consider rate during training. The framework indeed provides a scalable coding functionality, where RNN with convolutional and deconvolution layers is reported to perform decently. Toderici et al. then proposed an improved version by upgrading the encoder with a neural network similar to PixelRNN to compress the binary codes. The performance is reportedly better than JPEG on Kodak image dataset using MS-SSIM evaluation metric. Johnston et al. further improve the RNN- based solution by introducing hidden-state priming. In addition, an SSIM-weighted loss function is also designed, and spatially adaptive bitrates mechanism is enabled. They achieve better results than BPG on Kodak image dataset using MS-SSIM as evaluation metric. Covell et al. support spatially adaptive bitrates by training stop-code tolerant RNNs. Ballé et al. proposes a general framework for rate-distortion optimized image compression. The use multiary quantization to generate integer codes and consider the rate during training, i.e. the loss is the joint rate-distortion cost, which can be MSE or others. They add random noise to stimulate the quantization during training and use the differential entropy of the noisy codes as a proxy for the rate. They use generalized divisive normalization (GDN) as the network structure, which consists of a linear mapping followed by a nonlinear parametric normalization. The effectiveness of GDN on image coding is verified in an existing design. Ballé et al. then propose an improved version, where they use 3 convolutional layers each followed by a down-sampling layer and a GDN layer as the forward transform. Accordingly, they use 3 layers of inverse GDN each followed by an up-sampling layer and convolution layer to stimulate the inverse transform. In addition, an arithmetic coding method is devised to compress the integer codes. The performance is reportedly better than JPEG and JPEG 2000 on Kodak dataset in terms of MSE. Furthermore, Ballé et al. improve the method by devising a scale hyper-prior into the auto-encoder.They transform the latent representation ^^ with a subnet ℎ^ to ^^ = ℎ^(^^) and ^^ will be quantizedand transmitted as side information. Accordingly, the inverse transform is implemented with a subnet ℎ^attempting to decode from the quantized side information ^^^ to the standard deviation of the quantized ^^^, which will be further used during the arithmetic coding of ^^^. On the Kodak image set, their method is slightly worse than BGP in terms of PSNR. D. Minnen et al. further exploit the structures in the residue space by introducing an autoregressive model to estimate both the 11 F1254298PCTstandard deviation and the mean. In an existing design, Z. Cheng et al. use Gaussian mixture model to further remove redundancy in the residue. The reported performance is on par with VVC on the Kodak image set using PSNR as evaluation metric. 2.4 Neural networks for video compression Similar to conventional video coding technologies, neural image compression serves as the foundation of intra compression in neural network-based video compression, thus development of neural network-based video compression technology comes later than neural network-based image compression but needs far more efforts to solve the challenges due to its complexity. Starting from 2017, a few researchers have been working on neural network-based video compression schemes. Compared with image compression, video compression needs efficient methods to remove inter- picture redundancy. Inter-picture prediction is then a crucial step in these works. Motion estimation and compensation is widely adopted but is not implemented by trained neural networks until recently. Studies on neural network-based video compression can be divided into two categories according to the targeted scenarios: random access and the low-latency. In random access case, it requires the decoding can be started from any point of the sequence, typically divides the entire sequence into multiple individual segments and each segment can be decoded independently. In low-latency case, it aims at reducing decoding time thereby usually merely temporally previous frames can be used as reference frames to decode subsequent frames. 2.4.1 Low-latency Chen et al. are the first to propose a video compression scheme with trained neural networks. They first split the video sequence frames into blocks and each block will choose one from two available modes, either intra coding or inter coding. If intra coding is selected, there is an associated auto- encoder to compress the block. If inter coding is selected, motion estimation and compensation are performed with tradition methods and a trained neural network will be used for residue compression. The outputs of auto-encoders are directly quantized and coded by the Huffman method. Chen et al. propose another neural network-based video coding scheme with PixelMotionCNN. The frames are compressed in the temporal order, and each frame is split into blocks which are compressed in the raster scan order. Each frame will firstly be extrapolated with the preceding two reconstructed frames. When a block is to be compressed, the extrapolated frame along with the context of the current block are fed into the PixelMotionCNN to derive a latent representation. Then the residues are compressed by the variable rate image scheme. This scheme performs on par with H.264. 12 F1254298PCTLu et al. propose the real-sense end-to-end neural network-based video compression framework, in which all the modules are implemented with neural networks. The scheme accepts current frame and the prior reconstructed frame as inputs and optical flow will be derived with a pre-trained neural network as the motion information. The motion information will be warped with the reference frame followed by a neural network generating the motion compensated frame. The residues and the motion information are compressed with two separate neural auto-encoders. The whole framework is trained with a single rate-distortion loss function. It achieves better performance than H.264. Rippel et al. propose an advanced neural network-based video compression scheme. It inherits and extends traditional video coding schemes with neural networks with the following major features: 1) using only one auto-encoder to compress motion information and residues; 2) motion compensation with multiple frames and multiple optical flows; 3) an on-line state is learned and propagated through the following frames over time. This scheme achieves better performance in MS-SSIM than HEVC reference software. J. Lin et al. propose an extended end-to-end neural network-based video compression framework. In this solution, multiple frames are used as references. It is thereby able to provide more accurate prediction of current frame by using multiple reference frames and associated motion information. In addition, motion field prediction is deployed to remove motion redundancy along temporal channel. Postprocessing networks are also introduced in this work to remove reconstruction artifacts from previous processes. The performance is better than H.265 by a noticeable margin in terms of both PSNR and MS-SSIM. Eirikur et al. propose scale-space flow to replace commonly used optical flow by adding a scale parameter based on an existing framework. It is reportedly achieving better performance than H.264. Z. Hu et al. propose a multi-resolution representation for optical flows. Concretely, the motion estimation network produces multiple optical flows with different resolutions and let the network to learn which one to choose under the loss function. The performance is slightly improved and better than H.265. 2.4.2 Random access Wu et al. propose a neural network-based video compression scheme with frame interpolation. The key frames are first compressed with a neural image compressor and the remaining frames are compressed in a hierarchical order. They perform motion compensation in the perceptual domain, i.e. deriving the feature maps at multiple spatial scales of the original frame and using motion to warp the feature maps, which will be used for the image compressor. The method is reportedly on 13 F1254298PCTpar with H.264. Djelouah et al. propose a method for interpolation-based video compression, wherein the interpolation model combines motion information compression and image synthesis, and the same auto-encoder is used for image and residual. Amirhossein et al. propose a neural network-based video compression method based on variational auto-encoders with a deterministic encoder. Concretely, the model consists of an auto-encoder and an auto-regressive prior. Different from previous methods, this method accepts a group of pictures (GOP) as inputs and incorporates a 3D autoregressive prior by taking into account of the temporal correlation while coding the laten representations. It provides comparative performance as H.265. 2.5 Cool-chic video Neural compression methods usually train an auto-encoder-based structure, where an encoder network (also known as analysis transform) compress an image into a quantized latent, and a corresponding decoder network (also known as synthesis transform) reconstruct image from latent. During inference stage, trained encoder extracts latent from image and encoded it into a bitstream using entropy model. Here decoder and entropy model are general (hence usually large) and invariant at receiver side, while the latent is the only image-dependent component that needs to be transmitted from emitter to receiver. Fig. 3 illustrates an inter-picture decoding process of cool-chic video for low-delay application. ARM stands for auto-regressive probability model. Three levels of latent grids are illustrated. For the model shown in Fig.3, the example constraints for training may be defined as follows: Reconstruction: ^^^^^ = D(^^^௧ , ^^௧),Rate:Fig.3 presents the inter-picture decoding process of cool-chic video for low-delay application. Cool-chic video fits to each single image in a video, causing two differences compared with above method. First cool-chic video omits encoder (analysis transform) module and instead relies on optimization (backpropagation) to find the optimal latent and network parameters per image. Second, in addition to latents, network (i.e. entropy model, upsampling, and synthesis) parameters need to be transmitted as well since they are image-dependent now. Note in cool-chic video, removing the requirement for generality allow the networks to be very small, providing low decoding complexity. In the following, the decoding steps and encoding process are described. Network parameters. At first decoding step, an entropy coding engine named range coder parses from bitstream the network parameters, i.e. ^^ for ARM, ^^ for upsampling, ^^ for synthesis. The probability distribution of each parameter is modeled via a continuous Laplace distribution with non-parameterized mean and variance. 14 F1254298PCTLatent grids. Then given ^^, an auto-regressive probability model parameterized by the MLP ^^టdrives the entropy decoding of a hierarchical set of two-dimensional latent variables ^^^௧, where ^^ is time stamp. Specifically, the probability distribution of each latent element is modeled via a Laplace distribution as well, where the expectation and scale parameters of the distribution for each element are autoregressively predicted by ^^టfrom the causal neighborhood of that element.Here MLP ^^ట is a three-layer fully connected network of width < 20. By default latent ^^^ isrepresented with shape ofare height and widthof the image, L is the level of hierarchy and typically equals 7. Latent grids span a large range of resolutions to facilitate compact representation of structures at different frequencies. Upsampling. Next with ^^, a transposed convolutional layer ^^௨upsamples low-resolution latentgrids to yield a denser version. In particular latent grid of sizeis upsampled ^^each time by a factor of 2. The upsampled latents are concatenated along the channel dimensionto obtain a dense 3D representation of size ^^ × ^^ × ^^. Note a single upsampling kernel is appliedacross different latents and resolutions for simplicity.Synthesis. Now using decoded parameter ^^, a synthesis MLP ^^ఏ maps the ^^ × ^^ × ^^ latents tothe decoded features, i.e. the YUV residue channels ^^ , optical flowand element- weight map ^^. Here MLP ^^ఏis a three-layer convolutional network of width < 30 and kernel size= 1 × 1.Inter coding module. At last decoding step, non-parameterized inter coding module reconstructsYUV channels from synthesis outputs,where ⊙ denotes element-wise multiplication. Here inter prediction is generated from temporal context by warping reference signal ^^^ using optical flow. In case where interfails (e.g. when there are motion inconsistency or occlusion), the dense weight map ^^ ∈ ^0,1^ு×^is applied before adding the inter prediction with residue, to mask out the counterproductive prediction area. ^^ can be considered as intra-inter weight map. Encoding. Encoding frame ^^௧means overfitting the parameters {^^^௧, ^^, ^^, ^^}, to minimize the rate-distortion cost (minimization typically achieved with backpropagation-based optimization),i.e.{^^^௧, ^^,^^, ^^} = arg min^^൫^^ఏ൫^^௨(^^^௧)൯^^^௧^ + ^^ఏ൫^^௨(^^^௧)൯^^^^⊙ ^^^^^^^^൫^^^௧ି^, ^^ఏ൫^^௨(^^^௧)൯^^^௧ି^→௧^൯, ^^௧൯ − ^^^^^^^^ଶ^^ట(^^^௧)where ^^ is the original YUV channels, ^^^௧^, and ^^^^ denote slicing out ^^௧,^^ from the synthesis output. ^^ and −^^^^^^ଶ^^ట(^^^௧) stands for the distortion and rate term15 F1254298PCT(approximated with entropy). ^^ is the parameter to govern the rate-distortion trade-off of each picture and remains fixed for all-inter pictures in a video. 3 Problems The current overfitted neural codec has the following problems: 1. Many of these methods implicitly capture intra and inter-frame dependencies by processing multiple frames simultaneously. This forbids the usage of low-latency (a.k.a., low-delay) coding configurations since dozens / hundreds of frames are encoded together. 2. Motion is not sufficiently handled in overfitted neural codec. For example, cool-chic video only uses one reference picture in low-delay configuration. 3. Motion representation and residue representation are not decoupled. For example, cool-chic video encodes both residue and motion information into a set of latent representations. When large temporal variance occurs, this non-decoupled design may cause performance degradation. 4 Detailed Solutions The detailed solutions below should be considered as examples to explain general concepts. These solutions should not be interpreted in a narrow way. Furthermore, these solutions can be combined in any manner. The techniques described herein provide an overfitted neural compression method with decoupled residue and motion representation. The overall model architecture is shown in Fig. 4. Fig. 4 illustrates an inter-picture decoding process of proposed L3 for low-delay application. Three levels of latent grids are illustrated. Compared with Fig. 3, motion and residual representations are decoupled here, and multiple references are exploited. Besides, additional constraints may be added on the flow and prediction for training. For the model shown in Fig. 4, the example constraints for training may be defined as follows: Flow:Prediction:Reconstruction: ^^^^^ = D(^^^௧ , ^^௧),Motion Rate:Residue Rate:At a high level, some embodiments of the present disclosure introduce a dedicated motion module to decouple motion representation from residue representation. To foster generation of qualified prediction, some embodiments of the present disclosure introduce additional constraints to guide the learning of motion module. In summary, the proposed solutions include the following features: 1. It is proposed to decouple the representations for residue and motion signal as shown in 16 F1254298PCTFig.4. a. In one example, motion module mimics the logic of cool-chic, i.e. including an autoregressive probability model, hierarchical latents, learned upsampling, and a synthesis network. b. In one example, the resolution of latents in motion module is lower than its residueare the spatialsize of the largest latent grid in motion and residue modules, respectively. c. In one example, size of ARM ^^ట^ and synthesize network ^^ఏ^ is reduced compared with their counterpartsd. In one example, to bring back the resolution of motion synthesis output, a bilinear upsampling ^^^is inserted after the synthesis network. It is proposed to use multiple references to generate inter prediction. In this way, diversities of temporal contexts are increased. a. In one example, synthesis network in motion module outputs multiple optical flows between previously coded frames and current frame, i.e.,Then multiple predictions are obtained by warping the ^^௧^ reference ^^^௧ି^using=where ^^^ denotes the ^^௧^ prediction.b. In one example, each prediction is equipped with signal-dependent weights. i. In one example, signal-dependent weights are obtained by adding more output channels to the synthesis network in residual branch (similar to intra-inter weight map ^^ ). Denote dense signal-dependent weight maps as ^^௧ି^→௧ ∈^ , final inter prediction is generated as:⊙ ^^ wheredenotes number of active references and equals 4 by default. Note sum of all ^^ maps shall be 1,= 1. Hence residual synthesis network ^only generates first ^^ − 1 maps and derives the ^^௧^^^ map via above equation.ii. In one example, signal-dependent weights are obtained by adding more output channels to the synthesis network in motion branch (similar to intra-inter weight map ^^ ). Denote dense signal-dependent weight maps as ^^௧ି^→௧ ∈^ , final inter prediction is generated as:⊙ ^^ wheredenotes number of active references and equals 4 by default. Note sum of all ^^ maps shall be 1, i.e.= 1. Hence motion synthesis network ^only generates first ^^ − 1 maps and derives the ^^௧^^^ map via above equation.It is proposed to add additional constraints for training the neural codec. 17 F1254298PCTa. In one example, it is proposed to supervise the learning of motion module with knowledges of a pretrained mature model. In particular, the optical flowbetween the reference ^^^௧ି^and current picture ^^௧will be computed and serve as the groundtruth forThe following flow constraint is added into the distortion term at the initial training stage: ^^^^^௪i. In one example, the pretrained mature model is RAFT large model in PyTorch. ii. In one example, the pretrained mature model is RAFT small model in PyTorch. b. In one example, it is proposed to add a prediction constraint as below: ^^^^^ௗ =^^ (^^^, ^^௧),where ^^^ is the prediction derived in bullet 2.a.4. It is proposed to use different training objectives in different training stages or for different pictures. Denote optimal parameters to fit a picture in terms of RD cost as {^^^௧^, ^^^, ^^^, ^^^} and {^^^௧௭, ^^௭, ^^௭, ^^௭}. For simplicity, let ^^^and ^^௭represent the forward processes of residue and motion branches, i.e., ^^^ =Also, abbreviatea. In one example, constraints on flow, prediction, and reconstruction can be written as below, where the subscripts and superscripts of all summations are omitted to be concise. Note ^^ and ^^ indicate dependency on residue and motion parametersrespectively.=^^(^^^^^^௧^ + ^^^^^^^ ⊙ ∑^^^^^^^^ ⊙ ^^^^^^^^(^^^௧ି^ , ^^௭^^^^^), ^^௧)b. In one example, constraints on rate can be written as below:^^^^^^^ௗ௨^ = −^^^^^^ଶ^^ట^(^^^^௧ )c. In one example, a weighted combination of above constraints gives the overall training objective as below. Here, ^^^^^௪and ^^^^^ௗcontrol the contribution of additional constraints during training, whileis the parameter to govern the rate- distortion trade-off of ^^௧^ picture.{^^^^,^^^, ^ ^ ௭ ௭ ௭ ௭௧ ^^ , ^^ , ^^^௧ , ^^ , ^^ ,^^ }i. In one example, values of ^^^^^௪and ^^^^^ௗchange with training as below, 18 F1254298PCTwhere ^^^^^^ denotes current iteration number, ^^^^^௪and ^^^^^ௗare the number of iterations to impose flow and prediction constraints, respectively.ii. In one example, ^^௧is altered every 8 pictures as below, to ensure there are pictures of high quality for references at set intervals.5 Embodiment 5.1 Embodiment #1 L3: Low-complexity and Low Latency Neural Compression from a Single Video Brief Summary Most neural compression methods fit expressive models to large datasets, often resulting in considerable decoding complexity. Recently, there has been growing interest in codecs that overfit a single instance due to their lower decoding complexity. However, these codecs typically require access to the entire video or a significant number of frames simultaneously for training, making them unsuitable for applications that demand low latency (low-latency coding must avoid using future frames when encoding the current picture). To address this issue, some example embodiments of the present disclosure introduce L3, a low-complexity and low-latency neural compression approach. L3 builds on the cool-chic video model, with a decoding complexity of 1200 multiplications per pixel. Compared to the cool-chic video model, L3 significantly enhances the exploitation of temporal context by incorporating a well-designed, complexity-constrained motion component. Some example embodiments of the present disclosure also develop simple and effective training strategies to facilitate the learning of this motion component. With its improved motion capture capabilities, L3 significantly outperforms the cool-chic video model while maintaining the same level of complexity, achieving bitrate reduction on JVET test sequences. This may be one of the neural video codecs with the lowest decoding complexity (around 1k MAC / pixel) while achieving performance comparable to HEVC. Introduction Neural video compression (NVC) relies entirely on neural networks to establish its compression framework . This approach can be conceptualized as mapping raw video data to a latent representation (e.g., feature maps) and then reconstructing the original video from the latent representation. During training, the model is optimized to fit the training dataset. To ensure good 19 F1254298PCTgeneralization to unseen data, two conditions must be met: the training dataset must be sufficiently large to represent natural video distributions, and the model must have enough capacity to learn and capture patterns within this extensive data, which inevitably increases decoding complexity. While recent neural compression methods demonstrate strong rate-distortion (RD) performance, their high decoding complexity limits their applicability in scenarios with stringent constraints, particularly on mobile or edge devices with limited processing power. An alternative approach to reducing decoding complexity involves overfitting the neural codec to a single instance . In this method, the neural model is tailored to encode only a single video, which generally exhibits lower variance compared to a diverse set of natural videos. However, many of these methods implicitly capture intra and inter-frame dependencies by processing multiple frames simultaneously. This forbids the usage of low-latency (a.k.a low-delay) coding configurations since dozens / hundreds of frames are encoded together. Cool-chic video , an extension of the cool-chic image codec , illustrates an overfitting approach that supports low-latency compression. It processes each frame individually using overfitted parameters. Specifically, cool-chic video encodes both residue and motion information into a set of latent representations and decodes these from the latents using a synthesis module composed of simple convolutional layers. However, even within a single video, large temporal variance can occur, and the simplified design of cool-chic video may struggle with this variability, potentially leading to performance degradation. Details of cool-chic video will be discussed further later. Some example embodiments of the present disclosure introduce L3, a neural compression approach designed for low complexity and low latency. L3 builds upon the cool-chic video framework but significantly enhances its RD performance while maintaining the same decoding complexity. Specifically, inspired by the separate design of residual and motion coding in classical neural video compression, some example embodiments of the present disclosure propose a dedicated motion module distinct from the residual coding module. To manage complexity, the motion module operates at a resolution 4x lower than that of the residual module. Additionally, some example embodiments of the present disclosure have developed simple yet effective training strategies to improve the learning of the motion module. These advancements result in a BD-rate reduction compared to cool-chic video on the JVET test dataset . To the best of knowledge, L3 is one of the neural video codecs with the lowest decoding complexity (around 1k MAC / pixel) while achieving performance comparable to HEVC. Background: Cool-chic Video Neural compression methods usually train an auto-encoder-based structure, where an encoder network (also known as analysis transform) compress an image into a quantized latent, and a 20 F1254298PCTcorresponding decoder network (also known as synthesis transform) reconstruct image from latent. During inference stage, trained encoder extracts latent from image and encoded it into a bitstream using entropy model. Here decoder and entropy model are general (hence usually large) and invariant at receiver side, while the latent is the only image-dependent component that needs to be transmitted from emitter to receiver. Fig.3 presents the inter-picture decoding process of cool-chic video for low-delay application. Cool-chic video fits to each single image in a video, causing two differences compared with above method. First cool-chic video omits encoder (analysis transform) module and instead relies on optimization (backpropagation) to find the optimal latent and network parameters per image. Second, in addition to latents, network (i.e. entropy model, upsampling, and synthesis) parameters need to be transmitted as well since they are image-dependent now. Note in cool-chic video, removing the requirement for generality allow the networks to be very small, providing low decoding complexity. In the following, the decoding steps and encoding process are desribed. Network parameters. At first decoding step, an entropy coding engine named range coder parses from bitstream the network parameters, i.e. ^^ for ARM, ^^ for upsampling, ^^ for synthesis. The probability distribution of each parameter is modeled via a continuous Laplace distribution with non-parameterized mean and variance. Latent grids. Then given ^^, an auto-regressive probability model parameterized by the MLP ^^టdrives the entropy decoding of a hierarchical set of two-dimensional latent variables ^^^௧, where ^^ is time stamp. Specifically, the probability distribution of each latent element is modeled via a Laplace distribution as well, where the expectation and scale parameters of the distribution for each element are autoregressively predicted byfrom the causal neighborhood of thatHere is a three-layer fully connected network of width < 20. By default latent ^^^represented with shape ofare height and widthof the image, L is the level of hierarchy and typically equals 7. Latent grids span a large range of resolutions to facilitate compact representation of structures at different frequencies. Upsampling. Next with ^^, a transposed convolutional layer ^^௨upsamples low-resolution latentgrids to yield a denser version. In particular latent grid of sizeis upsampled ^^each time by a factor of 2. The upsampled latents are concatenated along the channel dimensionto obtain a dense 3D representation of size ^^ × ^^ × ^^. Note a single upsampling kernel is appliedacross different latents and resolutions for simplicity.Synthesis. Now using decoded parameter ^^, a synthesis MLP ^^ఏ maps the ^^ × ^^ × ^^ latents tothe decoded features, i.e. the YUV residue channels ^^௧, optical flow ^^௧ି^→௧, and element-wise 21 F1254298PCTweight map ^^. Here MLP ^^ఏis a three-layer convolutional network of width < 30 and kernel size= 1 × 1.Inter coding module. At last decoding step, non-parameterized inter coding module reconstructsYUV channels from synthesis outputs,^^^௧ = ^^௧ + ^^ ⊙ ^^^^^^^^(^^^௧ି^, ^^௧ି^→௧)where ⊙ denotes element-wise multiplication. Here inter prediction is generated from temporal context by warping reference signalusing optical flowIn case where inter predictionfails (e.g. when there are motion inconsistency or occlusion), the dense weight map ^^ ∈ [0,1^ு×^is applied before adding the inter prediction with residue, to mask out the counterproductive prediction area. ^^ can be considered as intra-inter weight map. Encoding. Encoding frame ^^௧means overfitting the parameters {^^^௧, ^^, ^^, ^^}, to minimize the rate-distortion cost (minimization typically achieved with backpropagation-based optimization),i.e.where ^^௧ is the original YUV channels, [^^௧^,and [^^^ denote slicing out ^^௧,and^^ from the synthesis output. ^^ and −^^^^^^ଶ^^ట(^^^௧) stands for the distortion and rate term(approximated with entropy). ^^ is the parameter to govern the rate-distortion trade-off of each picture and remains fixed for all-inter pictures in a video. L3: Improve Cool-chic Video The model and training improvements in L3 relative to cool-chic video is described in this section. The overall model architecture is shown in Fig.4. At a high level, some example embodiments of the present disclosure introduce a dedicated motion module to decouple motion representation from residue representation. To foster generation of qualified prediction, some example embodiments of the present disclosure introduce additional constraints to guide the learning of motion module. In the following, new features in L3 are described. Decoupled motion representation. In both conventional video coding such as HEVC and VVC and classical neural compression such as M-LVC and DCVC , motion signal is represented and coded separately from residual signal. For example, HEVC and VVC store motion information as block-wise motion vector while transmitting residue in dense pixel level. In M-LVC and DCVC, dedicated autoencoders are designed to handle motion information. Inspired by above design, some example embodiments of the present disclosure propose to decouple the representations for residue and motion signal as shown in Fig.4. Motion module mimics the logic of cool-chic, i.e. including an autoregressive probability model, 22 F1254298PCThierarchical latents, learned upsampling, and a synthesis network. Besides these commonalities, differences also exists to accommodate characteristics of motion signal. Motion signal typically exhibits lower variances than original / residual signal, hence keep the resolution of latents inmotion module lower than its residue counterpart:where (^^௭, ^^௭) and (^^^, ^^^) are the spatial size of the largest latent grid in motion and residue modules, respectively. Given the downscaled spatial resolution in motion representation, reduce size of ARM ^^ట^ and synthesize network ^^ఏ^ accordingly compared with their counterparts ^^ట^ and ^^ఏ^. Later to bring back the resolution, a bilinear upsampling ^^^is inserted after the synthesis network. Note that with 4 times lower resolution as well simplified ARM and synthesis designs, computational cost for representing motion drops significantly in comparison with representing residue. Multiple references. For increasing diversities of temporal contexts, some example embodiments of the present disclosure introduce multiple references in L3. To this end, it is proprosed to have synthesis network in motion module output multiple optical flows between previously coded frames and current frame, i.e.,Then, multiple predictions can be obtained,i.e. by warping the ^^௧^ reference ^^^௧ି^ using^^^ = ^^^^^^^^(^^^௧ି^ , ^^௧ି^→௧)where ^^^denotes the ^^௧^prediction. Actual contribution from each prediction may vary according to video contents, therefore it is necessary to equip these predictions with signal-dependent weights. To fully utilize existing infrastructures, a natural way is adding more output channels to the synthesis network either in residual branch or motion branch (similar to intra-inter weight map ^^). It is empirically found that residue branch is more suitable to play the role, which is likely because its pixel-level accuracyleads to more accurate weighting. Given dense weight maps∈ [0,1^ு×^ , final interprediction is generated as:where N denotes number of active references and equals 4 by default. Note sum of all ^^ maps shall be 1, i.e.Hence residual synthesis network ^^ ^ ௧^ఏ only generates first ^^ − 1 maps and derives the ^^ ^^ map23 F1254298PCTvia above equation. Additional constraints. It is noted that it not easy for a network to capture the motion when training from scratch, especially in case of large motion offset (e.g. on sequences of fast motion or high-resolution). To mitigate this issue, first consider supervise the learning of motion module with knowledges a pretrained mature model, namely RAFT . In particular, the optical flowbetween the reference ^^^௧ି^and current picture ^^௧will be computed and serve as the groundtruth The following flow constraint is added into the distortion term at the initial training stage,RAFT is a larger model than the motion module in this proposed method, hence above constraint helps distill the knowledge from RAFT into the motion module in this proposed method. To prioritize the prediction quality and reduce dependencies of the model on residual signals (representing residue information is more expensive in terms of bitrate from experience), some example embodiments of the present disclosure further propose a prediction constraint as below,where ^^^is the prediction from previous equation. This constraint has two implications. First, by explicitly pushing the warped signal to the groundtruth image, the flow module has a chance to generate optical flow of accuracy beyond the RAFT model. Second, try to maximize the capability of each prediction ^^^instead of the final one, i.e. ^^ from previous equation, to reduce unexpected bias toward a prediction only performing well at the beginning. Optimization with additional constraints. This part describes the encoding process of L3, i.e. looking for optimal parameters {^^^௧^, ^^^, ^^^, ^^^} and {^^^௧௭, ^^௭, ^^௭, ^^௭} to fit a picture in terms of RD cost. For simplicity, let ^^^and ^^௭represent the forward processes of residue and motionbranches, i.e., ^^ ௭^^^௭ = ^^^ ^^^ఏ^൫^^௨^(^^^௧ )൯^. Also, abbreviate ^^௧ି^→௧ andto ^^^and ^^^.Constraints on flow, prediction, and reconstruction can be written as:^^௭^^^ௗ = ∑^^(^^^, ^^௧) = ∑^^(^^^^^^^^(^^^௧ି^ , ^^௭[^^^^), ^^௧)=^^(^^^[^^௧^ + ^^^[^^^ ⊙ ∑^^^[^^^^ ⊙ ^^^^^^^^(^^^௧ି^ , ^^௭[^^^^), ^^௧)where the subscripts and superscripts of all summations are omitted to be concise. Note ^^ and ^^ 24 F1254298PCTindicate dependency on residue and motion parameters respectively. Constraints on rate can bewritten as:A weighted combination of above constraints gives the overall training objective,{^^^^,^^^, ^^^, ^ ^ ௭ ௭ ௭ ௭௧ ^ , ^^^௧ , ^^ , ^^ ,^^ }Here, ^^^^^௪and ^^^^^ௗcontrol the contribution of additional constraints during training, while ^^௧is the parameter to govern the rate-distortion trade-off of ^^௧^picture. Values of ^^^^^௪and ^^^^^ௗchange with training as:where ^^^^^^ denotes current iteration number, ^^^^^௪and ^^^^^ௗare the number of iterations to impose flow and prediction constraints, respectively.is altered every 8 pictures as below, to ensure there are pictures of high quality for references at set intervals. Note ^^ remains fixed in previousequation for all inter pictures in a video.5.2 Embodiment #2 CD: Cool-chic Video with Decoupled Representation Brief Summary Most neural compression methods fit expressive models to large datasets, often resulting in considerable decoding complexity. Recently, there has been growing interest in codecs that overfit a single instance due to their reduced decoding complexity. However, these codecs typically require access to the entire video or a significant number of frames simultaneously for encoding, which limits their practicality, particularly in low-delay coding configurations. To address this issue, some example embodiments of the present disclosure introduce CD, a low-complexity neural video compression method with a decoding complexity of 1.2k MACs per decoded pixel. CD builds upon the cool-chic video model by incorporating decoupled representations for motion and residue. This approach significantly enhances the exploitation of temporal context through a well-designed, complexity- constrained motion component. some example embodiments of the present disclosure also introduce simple yet effective training strategies to facilitate the learning of this motion 25 F1254298PCTcomponent. With its improved motion capture capabilities, CD substantially outperforms the cool-chic video model while maintaining the same level of complexity, achieving up to a 56% and on average 28% bitrate reduction on the HEVC test dataset. The picture-wise overfitting nature of CD offers flexibility for various encoding configurations, including support for low-delay coding. To the best of knowledge, CD is one of the neural video codecs with the lowest decoding complexity (around 1k MACs / pixel) while achieving performance comparable to HEVC. Introduction Neural video compression relies entirely on neural networks to establish its compression framework . This approach can be conceptualized as mapping raw video data to a latent representation (e.g., feature maps) and then reconstructing the original video from this latent representation. During training, the model is optimized to fit the training dataset. To ensure good generalization to unseen data, two key conditions shall be met: the training dataset must be sufficiently large to approach natural video distributions, and the model must have enough capacity to learn and capture patterns within this extensive data. However this inevitably increases decoding complexity. While recent neural compression methods demonstrate strong rate-distortion (RD) performance, their high decoding complexity limits their applicability in scenarios with stringent constraints, particularly on mobile or edge devices with limited processing power. An alternative approach to reducing decoding complexity involves overfitting the neural codec to a single instance . In this method, the neural model is tailored to encode only a single video, which generally exhibits lower variance compared to a diverse set of natural videos. However, many of these methods have very coarse overfitting granularity, making them less flexible. For example, NeRV-based methods typically encode hundreds of frames together, which prevents their usage in low-delay coding configurations. Cool-chic video , an extension of the cool-chic image codec , illustrates an overfitting approach that supports picture-wise encoding. It processes each frame individually using overfitted parameters. Specifically, cool-chic video encodes both residue and motion information into a set of latent representations and reconstructs these using a synthesis module composed of simple convolutional layers. However, even within a single video, significant temporal variance can occur, and the simplified design of cool-chic video may struggle with this variability, potentially leading to performance degradation. Further details of cool-chic video will be discussed later. Some example embodiments of the present disclosure introduce CD, a neural video compression method with picture-wise overfitting. It is named as CD as it builds upon Cool-chic video, but with Decoupled representations for motion and residue. CD significantly enhances the RD performance 26 F1254298PCTof cool-chic video while maintaining the same level of decoding complexity. Specifically, inspired by the separate design of residual and motion coding in classical neural video compression , some example embodiments of the present disclosure propose a dedicated motion module distinct from the residual coding module. To manage complexity, the motion module operates at a resolution four times lower than that of the residual module. Additionally, some example embodiments of the present disclosure have developed simple yet effective training strategies to improve the learning of the motion module. These advancements result in an average 28% BD-rate reduction compared to cool-chic video on the HEVC test dataset . The picture-wise overfitting nature of CD offers flexibility for different encoding configurations, including support for low-delay coding. To the best of knowledge, CD is one of the neural video codecs with the lowest decoding complexity (around 1k MACs / pixel) while achieving performance comparable to HEVC. Background: Cool-chic Video Neural compression methods typically employ an auto-encoder-based structure, where an encoder network (analysis transform) compresses an image into a quantized latent representation, and a corresponding decoder network (synthesis transform) reconstructs image from the latent representation. During inference, the trained encoder extracts the latent representation from the image and encodes it into a bitstream using an entropy model. In this setup, the decoder and entropy model are general (hence usually large) and invariant at the receiver side, while the latent representation is the only image-dependent component that needs to be transmitted from the emitter to the receiver. Fig.3 presents the inter-picture decoding process of cool-chic video in low-delay configurations. Cool-chic video fits each individual image in a video, leading to two key differences compared to the conventional neural compression method. First, cool-chic video omits the encoder (analysis transform) module and instead relies on optimization (backpropagation) to find the optimal latent representation and network parameters for each image. Second, in addition to latents, the network parameters (i.e., entropy model, upsampling, and synthesis) must also be transmitted, as they are now image-dependent. In cool-chic video, removing the requirement for generality allow the networks to be very small, providing low decoding complexity. The following parts detail the decoding steps and encoding process. Network parameters. At the first decoding step, an entropy coding engine named range coder parses the bitstream to retrieve the network parameters, namely ^^ for the ARM, ^^ for upsampling, and ^^ for synthesis. The probability distribution of each parameter is modeled via a continuous Laplace distribution with non-parameterized mean and variance.Latent grids. Given ^^ , an auto-regressive probability model parameterized by the MLP ^^ట27 F1254298PCTcontrols the entropy decoding of a hierarchical set of two-dimensional latent variables ^^^௧, where ^^ denotes the timestamp. Specifically, the probability distribution of each latent element is modeled using a Laplace distribution. The expectation and scale parameters of this distribution are autoregressively predicted by ^^టbased on the causal neighborhood of the element. The MLP ^^టis a three-layer fully connected network with a width of less than 20. By default, the latentrepresentation ^^^ is organized in in a hierarchical grid structure with shapes...where ^^ and ^^ represent the height and width of the image, and ^^, thelevel, typically equals 7. These latent grids span a broad range of resolutions, enabling the compact representation of structures at different frequency levels.Upsampling. With ^^ , a transposed convolutional layer ^^௨ is employed to upsample low-resolution latent grids, producing a denser representation. Specifically, the latent grid of sizetimes, each time by a factor of 2. The upsampled latents areconcatenated along the channel dimension, resulting in a dense 3D representation of size^^ × ^^ × ^^. For simplicity, a single upsampling kernel is applied uniformly across different latentsand resolutions.Synthesis. Using the decoded parameter ^^, a synthesis MLP ^^ఏ maps the ^^ × ^^ × ^^ latents tothe decoded features, specifically the YUV residue channels ^^௧, optical flowand element- wise weight map ^^. The MLP ^^ఏis a three-layer convolutional network with a width of less than30 and a kernel size of 1 × 1.Inter coding module. In the final decoding step, a non-parameterized inter coding modulereconstructs the YUV channels from the synthesis outputs,^^^௧ = ^^௧ + ^^ ⊙ ^^^^^^^^(^^^௧ି^, ^^௧ି^→௧)where ⊙ denotes element-wise multiplication. Here inter prediction is generated from temporal context by warping the reference signal ^^^ using the optical flowIn case where prediction is inaccuratesuch as in scenarios with motion inconsistency or occlusionthe denseweight map ^^ ∈ [0,1^ு×^ is applied to mask out counterproductive areas of the prediction beforeadding it to the residue. The weight map ^^ can be considered as an intra-inter weight map. Encoding. Encoding frame ^^௧involves overfitting the parameters {^^^௧, ^^, ^^, ^^} to minimize therate-distortion cost, i.e.,{^^^௧ , ^^,^^, ^^} = arg min^^൫^^ఏ൫^^௨(^^^௧)൯[^^௧^ + ^^ఏ൫^^௨(^^^௧)൯[^^^where ^^ represents the original YUV channels, while [^^௧^,and [^^^ denote the corresponding to ^^௧, ^^௧ି^→௧, and ^^ from the synthesis output. The ^^ refers to the distortion, and 28 F1254298PCT−^^^^^^ଶ^^ట(^^^௧) stands for the rate term approximated by entropy. The parameter ^^ governs the rate- distortion trade-off for each picture and remains fixed across all-inter pictures in a video. This minimization is typically achieved through backpropagation-based optimization. CD: Improve Cool-chic Video In this section, the model and training improvements in CD relative to cool-chic video is described. The overall model architecture is illustrated in Fig.4. At a high level, some example embodiments of the present disclosure introduce a dedicated motion module to decouple motion representation from residue representation. To ensure the generation of accurate predictions, some example embodiments of the present disclosure incorporate additional constraints to guide the learning of the motion module. In the following, the new features introduced in CD are described. Decoupled representations. In both conventional video coding, such as HEVC and VVC , and classical neural compression methods like M-LVC and DCVC , the motion signal is represented and coded separately from the residual signal. For instance, HEVC and VVC store motion information as block-wise motion vectors while transmitting the residue at a dense pixel level. In M-LVC and DCVC, dedicated autoencoders are designed to handle motion information. Inspired by these designs, some example embodiments of the present disclosure propose to decouple the representations for the residue and motion signals, as depicted in Fig.4. The motion module in the model according to some embodiments of the present disclosure follows the logic of cool-chic, incorporating an autoregressive probability model, hierarchical latents, learned upsampling, and a synthesis network. However, there are key differences to accommodate the specific characteristics of motion signals. Since motion signals typically exhibit lower variance than the original or residual signals, maintain a lower resolution for the latents in the motionmodule compared to its residual counterpart:where (^^௭, ^^௭) and (^^^, ^^^) are the spatial sizes of the largest latent grid in the motion and residual modules, respectively. Given the downscaled spatial resolution in the motion representation, reduce the size of the ARM ^^ట^and the synthesize network ^^ఏ^accordingly,compared to their counterpartsand ^^ఏ^. To restore the resolution, a bilinear upsampling layer^^^ is inserted after the synthesis network. By utilizing a 4× lower resolution and simplified ARMand synthesis designs, the computational cost for representing motion is significantly reduced compared to representing the residue. Multiple references. To increase the diversity of temporal contexts, some example embodiments of the present disclosure introduce multiple references in CD. Specifically, the synthesis network 29 F1254298PCTwithin the motion module outputs multiple optical flows between previously coded frames and the current frame, denoted asThese flows allow for generating multiple predictions bywarping the ^^௧^ reference ^^^௧ି^ using^^^ = ^^^^^^^^(^^^௧ି^ , ^^௧ି^→௧)where ^^^denotes the ^^௧^prediction. Since the actual contribution from each prediction may vary depending on the video content, it is essential to apply signal-dependent weights to these predictions. A natural way to achieve this is by adding additional output channels to the synthesis network, either in the residual branch or the motion branch (similar to the intra-inter weight map ^^). Empirical results indicate that the residue branch is more suitable for the task, likely due to its pixel-level accuracy, which leads to moreprecise weighting. Given dense weight maps∈ [0,1^ு×^ , the final inter prediction iscomputed as:where N denotes the number of active references and is set to 4 by default. The sum of all ^^ maps is constrained to equal 1,Thus, the residual synthesis network ^^ఏ^ generates the first ^^ − 1 maps, while the ^^௧^^^ map isderived from above equation. Additional constraints. It is noted that it is challenging for a network to capture motion accurately when training from scratch, particularly in cases of large motion offsets, such as in sequences with fast motion or high resolution. To address this issue, first consider supervising the learning of the motion module using the knowledges of a pretrained, mature model, i.e., RAFT . Specifically, the optical flowbetween the reference ^^^௧ି^and the current frame ^^௧is computed and used as the groundtruth for ^^௧ି^→௧. The following flow constraint is added to the distortion term during the initial training stage,where RAFT, being a larger model than the proposed motion module, distills its knowledge into the proposed model through this constraint. To prioritize prediction quality and reduce the dependency on residual signals, since representing residue information is more expensive in terms of bitrate, some example embodiments of the 30 F1254298PCTpresent disclosure introduce an additional prediction constraint,where ^^^is the prediction obtained from above equation. This constraint serves two purposes. First, by explicitly encouraging the warped signal to match the groundtruth image, the flow module 5 has the opportunity to generate optical flows with accuracy surpassing that of the RAFT model. Second, it aims to maximize the effectiveness of each individual prediction ^^^, rather than the final combined prediction ^^ from above equation. This helps prevent the model from developing an unintended bias toward predictions that perform well only in the early stages of training. Optimization with additional constraints. This part describes the encoding process of CD, 10 which involves finding the optimal parameters ^^^, ^^^} an௭ ௭ ௭ ௭d {^^^௧, ^^ , ^^ , ^^ } to fit a picture in terms of rate-distortion cost. For simplicity, let ^^^and ^^௭represent the forward processes of the residue and motion branches, respectively, i.e., ^^^^^ ^^^ఏ^൫^^௨^(^^^௭௧ )൯^. AbbreviateThe constraints on flow, prediction, and reconstruction can be expressed as:15^^^,௭^^^ = ^^(^^^௧ , ^^௧) = ^^(^^௧ + ^^ ⊙ ^^, ^^௧)= ^^(^^^[^^௧^ + ^^^[^^^ ⊙ ∑^^^[^^^^ ⊙ ^^^^^^^^(^^^௧ି^ , ^^௭[^^^^), ^^௧)where the subscripts and superscripts of all summations are omitted for brevity. Note that ^^ and20 ^^ indicate dependencies on residue and motion parameters, respectively. The constraints on therate can be written as follows.Aweighted combination of the above constraints forms the overall training objective,25 {^^^^,^^^, ^^^, ^^^ ௭ ௭ ௭ ௭௧ , ^^^௧ , ^^ , ^^ ,^^ }where ^^^^^௪ and ^^^^^ௗ control the contribution of the additional constraints during training, andgoverns the rate-distortion trade-off for the ^^௧^ picture. The values of ^^^^^௪ and ^^^^^ௗduring training as follows,3031 F1254298PCTwhere ^^^^^^ denotes the current iteration number, and ^^^^^௪and ^^^^^ௗare the numbers of iterations during which the flow and prediction constraints are imposed, respectively. The parameter ^^௧is adjusted every 8 pictures to ensure that some pictures are of high quality for use as references atset intervals.

[0030] More details of the embodiments of the present disclosure will be described below which are related to neural network-based video coding. The embodiments of the present disclosure should be considered as examples to explain the general concepts and should not be interpreted in a narrow way. Furthermore, these embodiments can be applied individually or combined in any manner.

[0031] As used herein, the term “video unit” may represent a picture of a video or a portion of the picture, such as, a coding tree block (CTB), a coding tree unit (CTU), a coding block (CB), a coding unit (CU), a prediction unit (PU), a transform unit (TU), a prediction block (PB), a transform block (TB), a subblock, a tile, a slice, a subpicture, a video processing unit comprising multiple samples / pixels, and / or the like. A video unit may be rectangular or non- rectangular.

[0032] Fig. 5 illustrates a flowchart of a method 500 for video processing in accordance with some embodiments of the present disclosure. The method 500 may be implemented during a conversion between an a current video unit of a video and a bitstream of the video with a neural network (NN)-based model comprising a first module and a second module. As used herein, an NN-based model may be a model based on neural network technologies. For example, an NN-based model may specify sequence of neural network modules (also called architecture) and model parameters. The neural network module may comprise a set of neural network layers. Each neural network layer specifies a tensor operation which receives and outputs tensor, and each layer has trainable parameters. It should be understood that the possible implementations of the NN-based model described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any way.

[0033] As shown in Fig.5, the method 500 starts at 502 where a first latent representation associated with motion information of the current video unit is determined by the first module with a first plurality of parameters. By way of example rather than limitation, the first module may comprise: an autoregressive probability model configured to generate a first set of latent representations by using a first set of parameters among the first plurality 32 F1254298PCTof parameters; an upsampling model configured to generate the first latent representation based on the first set of latent representations by using a second set of parameters among the first plurality of parameters; and a synthesis network configured to generate the motion information based on the first latent representation by using a third set of parameters among the first plurality of parameters.

[0034] With reference to Fig.4, the first module may correspond to the module 410. The first set of parameters is represented as ^^௭, and the processing corresponding to the autoregressive probability model with first set of parameters is represented as ^^ట^. The first set of latent representations are represented as latents ^^^௧௭. The second set of parameters is represented as ^^௭, and the processing corresponding to the upsampling model with second set of parameters is represented as ^^௨^. The third set of parameters is represented as ^^௭, and the processing corresponding to the synthesis network with the third set of parameters is represented as ^^ఏ^. It should be understood that the possible implementations of the first module described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any way. For example, the first module may further comprise one or more convolutional layers, downsampling layers, and / or the like.

[0035] At 504, a second latent representation associated with residual information of the current video unit is determined by the second module with a second plurality of parameters. The second module is independent from the first module. In other words, the second module is decoupled from the first module. By way of example, the module 410 and module 420 in Fig.4 are decoupled from each other.

[0036] By way of example rather than limitation, the second module may comprise: an autoregressive probability model configured to generate a second set of latent representations by using a fourth set of parameters among the second plurality of parameters, an upsampling model configured to generate the second latent representation based on the second set of latent representations by using a fifth set of parameters among the second plurality of parameters, and a synthesis network configured to generate the residual information based on the second latent representation by using a sixth set of parameters among the second plurality of parameters.

[0037] With reference to Fig. 4, the second module may correspond to the module 410. The fourth set of parameters is represented as ^^^, and the processing corresponding to the autoregressive probability model with fourth set of parameters is represented as ^^ట^. The second set of latent representations are represented as latents ^^^௧^. The fifth set of parameters is represented as ^^^, and the processing corresponding to the upsampling model with fifth 33 F1254298PCTset of parameters is represented as ^^௨^. The sixth set of parameters is represented as ^^^, and the processing corresponding to the synthesis network with the sixth set of parameters is represented as ^^ఏ^. It should be understood that the possible implementations of the second module described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any way. For example, the second module may further comprise one or more convolutional layers, downsampling layers, and / or the like.

[0038] In addition, the first plurality of parameters and the second plurality of parameters are indicated in the bitstream. For example, at an encoder side, the first plurality of parameters and the second plurality of parameters may be obtained by training the first module and the second module. This will be descried in details below. Furthermore, the first plurality of parameters and the second plurality of parameters may be encoded into the bitstream. At a decoder side, the first plurality of parameters and the second plurality of parameters may be decoded from the bitstream.

[0039] At 506, the conversion is performed based on the first latent representation and the second latent representation. In some embodiments, the conversion may include encoding the current video unit into the bitstream. Alternatively or additionally, the conversion may include decoding the current video unit from the bitstream. It should be understood that the above illustrations and / or examples are described merely for purpose of description. The scope of the present disclosure is not limited in this respect.

[0040] In view of the above, the first module for generating the first latent representation associated with motion information is independent from the second module for generating the second latent representation associated with residual information of the current video unit. Compared with the conventional solution, the proposed method can advantageously decouple the first latent representation and the second latent representation. Thereby, the proposed method can advantageously improve the coding performance.

[0041] In some embodiments, the first set of latent representations may only comprise one latent representation. Alternatively, the first set of latent representations may comprise more than one latent representations. For example, the more than one latent representations may have different resolutions, i.e., spatial sizes. In the example shown in Fig. 4, the latents ^^^௧௭comprise three latents, i.e., one latent with a resolution, one latent with aresolution of× ೈ^, and one lat ௭ ௭మ మ ent with a resolution of ^^ × ^^ .

[0042] In some embodiments, the second set of latent representations may only comprise one latent representation. Alternatively, the second set of latent representations may 34 F1254298PCTcomprise more than one latent representations. For example, the more than one latent representations may have different resolutions, i.e., spatial sizes. In the example shown inFig. 4, the latents ^^^^ comprise three late௧nts, i.e., one latent with a resolution ofర ,latent with a resolution of× ೈ^one , and o ^ ^మ మ ne latent with a resolution of ^^ × ^^ .

[0043] In some embodiments, a resolution of a latent representation having a largest resolution among the first set of latent representations may be lower than a resolution of a latent representation having a largest resolution among the second set of latentrepresentations. In the example shown in Fig. 4, the resolution ^^௭ × ^^௭ may be lower thanthe resolution ^^^ × ^^^. This is based on the observation that there is lower variance inmotion signals, and thus a size of the latent representation(s) associated with motion information may be reduced, so as to improve the coding efficiency.

[0044] In one example embodiment, the resolution of the latent representation having the largest resolution among the second set of latent representations may be an integer multiple (such as, double, triple or the like) of the resolution of the latent representation having the largest resolution among the first set of latent representations.

[0045] In some additional embodiments, a size of the autoregressive probability model of the first module may be smaller than a size of the autoregressive probability model of the second module. Additionally or alternatively, a size of the synthesis network of the first module may be smaller than a size of the synthesis network of the second module.

[0046] In some embodiments, the first module further may comprise a bilinear upsampling model configured to upsample an output of the synthesis network of the first module. A resolution of an output of the bilinear upsampling may be same as a resolution of an output of the synthesis network of the second module. With reference to Fig. 4, the processing corresponding to the bilinear upsampling model is represented as ^^^.

[0047] In some embodiments, a plurality of reference video units may be employed to generate a prediction for the current video unit, e.g., in a low-delay configuration. The low- delay configuration specifies that the pictures are coded in display order, so that the resulting structural encoding-decoding delay is suitable for low-delay communication applications.

[0048] In this case, the motion information of the current video unit may comprise respective optical flows between the current video unit and a plurality of reference video units of the current video unit. At 506, the respective optical flows may be generated based on the first latent representation. Moreover, a plurality of predictions for the current video unit may be determined based on the plurality of reference video units and the respective 35 F1254298PCToptical flows. By way of example, the i-th prediction (denoted as ^^^) among the plurality of predictions may be obtained by warping a reconstruction of the ^^௧^reference video unit (denoted as ^^^௧ି^) using a corresponding optical flowThe optical flowis measured between the ^^௧^reference video unit and the current video unit. For example, the warping operation may be implemented by shifting pixels in a video unit according to the motion information (such as motion vector or the like) represented by the optical flow.

[0049] In addition, a target prediction for the current video unit may be determined based on the plurality of predictions. In some embodiments, the target prediction may be determined based on a weighted sum of the plurality of predictions. For example, a plurality of weight maps for determining the weighted sum may be dependent on predicted signals within the plurality of predictions. In one example embodiment, at least a part of the plurality of weight maps may be generated by the first module. Additionally or alternatively, at least a part of the plurality of weight maps may be generated by the second module.

[0050] In some embodiments, (N-1) weight maps among the plurality of weight maps may be generated by the first module or the second module, where N represents a number of the plurality of weight maps. Moreover, a remaining one of the plurality of weight maps may be determined based on the (N-1) weight maps.

[0051] In some embodiments, in a training process, at least one of the following may be determined by training the first module and the second module based on a rate-distortion (RD) cost metric for the current video unit: the first set of latent representations, the first set of parameters, the second set of parameters, the third set of parameters, the second set of latent representations, the fourth set of parameters, the fifth set of parameters, or the sixth set of parameters.

[0052] In some embodiments, the RD cost metric may comprise an optical flow distortion term indicating a distortion between at least one optical flow generated by the first module and a ground truth for the at least one optical flow. The at least one optical flow may be measured between the current video unit and at least one reference video unit of the current video unit. In addition or alternatively, the RD cost metric may comprise a prediction distortion term indicating a distortion between at least one prediction for the current video unit and the current video unit. The at least one prediction may be determined based on the at least one optical flow generated by the first module.

[0053] In some embodiments, the optical flow distortion term may be determined asfollows:^^ = ∑^ୀே ^^ ൫ ோ^ி்^^^௪ ^ୀ^ ^^௧ି^→௧ ,^^௧ି^→௧൯,36 F1254298PCTwhere ^^^^^௪represents the optical flow distortion term, N represents a number of the at least one reference video unit,represents an optical flow between an i-th reference video unit and the current video unit that is generated by the first module,represents aground truth for the optical flow, andrepresents a distortion betweenIt should be understood that the optical flow distortion term may also bedetermined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0054] In some embodiments, the prediction distortion term may be determined as follows:where ^^^^^ௗrepresents the prediction distortion term, N represents a number of the at least one reference video unit, ^^^represents a prediction for the current video unit determined based on an optical flow between an i-th reference video unit and the current video unit thatis generated by the first module, ^^௧ represents the current video unit, and ^^(^^^ , ^^௧) representsa distortion between ^^^and ^^௧. It should be understood that the prediction distortion term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0055] In some embodiments, the RD cost metric may comprise: (i) a first distortion term indicating a distortion associated with at least one optical flow generated by the first module, (ii) a second distortion term indicating a distortion associated with at least one prediction for the current video unit, (iii) a third distortion term indicating a distortion associated with a reconstruction for the current video unit, (iv) a first rate term indicating a rate for data associated with the first module, (v) a second rate term indicating a rate for data associated with the second module, and / or the like.

[0056] In one example embodiment, the distortion associated with the at least one optical flow may be a distortion between the at least one optical flow and a ground truth for the at least one optical flow. The at least one optical flow may be measured between the current video unit and at least one reference video unit of the current video unit. In addition or alternatively, the distortion associated with the at least one prediction may be a distortion between the at least one prediction and the current video unit. The at least one prediction may be determined based on the at least one optical flow generated by the first module. In a further example embodiment, the distortion associated with the reconstruction may be a distortion between the reconstruction and the current video unit. The reconstruction may be determined based on the residual information and the at least one prediction. In a further example embodiment, the data associated with the first module comprise the first set of latent 37 F1254298PCTrepresentations. In a further example embodiment, the data associated with the second module comprise the second set of latent representations.

[0057] In some embodiments, the first distortion term may be determined as follows:where ^^^௭^^௪represents the first distortion term, ^^௭[^^^^ represents an optical flow between ani-th reference video unit and the current video unit that is generated by the first module,^^ோ^ி் represents a ground truth for the optical flow, ^^൫ [ ^ ோ^ி்^ ^^௭ ^^^ , ^^^ ൯ represents a distortionbetween ^^௭[^^^^ and ^^^ோ^ி், ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit. It should be understood that the first distortion term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0058] In some embodiments, the ground truth for the at least one optical flow may be generated by a pretrained model. For example, the pretrained model may be based on recurrent all-pairs field transforms (RAFT). In one example, the pretrained model may be RAFT large model in PyTorch. Alternatively, the pretrained model may be RAFT small model in PyTorch.

[0059] Thereby, training of the first module can be supervised by using the knowledges of the pretrained mature model, so as to capture motion more accurately.

[0060] In some embodiments, the second distortion term may be determined as follows:^^௭^^^ௗ = ∑^^(^^^^^^^^(^^^௧ି^ ,^^௭[^^^^), ^^௧),where ^^^௭^^ௗrepresents the second distortion term, ^^^௧ି^represents a reconstruction for an i- th reference video unit, ^^௭[^^^^ represents an optical flow between an i-th reference video unit and the current video unit that is generated by the first module, ^^^^^^^^(^^^an i-th prediction for the current video unit obtained by warping ^^^௧ି^ using ^^௭[^^^^ , ^^௧represents the current video unit, ^^(^^,^^) represents a distortion between A and B, and ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit. It should be understood that the second distortion term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0061] In some embodiments, the third distortion term may be determined as follows:where ^^^,௭^^^represents the third distortion term, ^^^[^^௧^ represents the residual information generated by the second module, ^^^[^^^represents an intra-inter weight map generated by thesecond module, ⊙ represents an element-wise multiplication, ^^^௧ି^ represents a38 F1254298PCTreconstruction for an i-th reference video unit, ^^௭[^^^^ represents an optical flow between ani-th reference video unit and the current video unit that is generated by the first module,^^^^^^^^(^^^௧ି^ ,^^௭[^^^^)represents an i-th prediction for the current video unit obtained by warping^^^௧ି^ using ^^௭[^^^^,represents a weight map for the i-th prediction generated by thesecond module, ^^௧represents the current video unit, ^^(^^,^^) represents a distortion between A and B, and ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit. It should be understood that the third distortion term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0062] In some embodiments, the first rate term may be determined as follows:where ^^^௭^௧^^^represents the first rate term, ^^^௧௭represents the first set of latent representations, and௭represents a distribution of ^^^௧. It should be understood that the first rate term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0063] In some embodiments, the second rate term may be determined as follows:where ^^^^^^^ௗ௨^represents the second rate term, ^^^௧^represents the second set of latent representations, and ^^ట^(^^^௧^)represents a distribution of ^^^௧^. It should be understood that the second rate term may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0064] In some embodiments, the current video unit may be comprised in a t-th picture ofthe video, and the RD cost metric for the current video unit may be defined as follows:where ^^^^^^ represents the RD cost metric, ^^^௭^^௪represents the first distortion term, ^^^௭^^ௗrepresents the second distortion term, ^^^,௭^^^represents the third distortion term, ^^^௭^௧^^^represents the first rate term, ^^^^^^^ௗ௨^represents the second rate term, ^^^^^௪represents a weight for ^^^௭^^௪, ^^^^^ௗrepresents a weight for ^^^௭^^ௗ,represents a parameter govern a rate-distortion trade-off for the t-th picture. It should be understood that the RD cost metric may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0065] In one example embodiment, different RD cost metrics may be applied for different training stages of the training process. By way of example rather than limitation, the weights 39 F1254298PCT^^^^^௪and ^^^^^ௗmay be determined as follows: 0.1, 0 if ^^^^^^ ∈ [0, ^^^^^௪)0.1 if ^^^^^^ ∈ [^^^^^௪, ^^^^^௪ + ^^^^^ௗ)0, 0 if ^^^^^^ ∈ [^^^^^௪ + ^^^^^ௗ , ^^௧^௧^^)where ^^^^^^ represents a current training iteration number during the training process, ^^^^^௪represents a number of training iterations in which a flow constraint may be imposed, ^^^^^ௗrepresents a number of training iterations in which a prediction constraint may be imposed, and ^^௧^௧^^represents a total number of training iterations performed in the training process. Thereby, the training objectives may be varied during the training process, such as to achieve better coding quality. It should be understood that the weights ^^^^^௪and ^^^^^ௗmay also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0066] In an additional or alternative example embodiment, different RD cost metrics may be applied for different pictures of the video. By way of example rather than limitation, theparameter may be determined as follows:where ^^^^^^ represents a base value, ^^ represents an adjusting factor unequal to 1, ^^^^^^^^represents a positive integer, and % represents a modulo operation. For example, theadjusting factor ^^ may be equal to 1.7, 2, or the like. ^^^^^^^^ may be equal to 8, 10, or the like. Thereby, it can be ensured that there are pictures of high quality at preset intervals, so that these high-quality pictures can be used as references for coding subsequent pictures of the video. Thereby, the coding quality can be improved. It should be understood that the parameter may also be determined in any other suitable manner. The scope of the present disclosure is not limited in this respect.

[0067] In view of the above, the solutions in accordance with some embodiments of the present disclosure can advantageously improve coding efficiency and coding quality.

[0068] According to further embodiments of the present disclosure, a non-transitory computer-readable recording medium is provided. The non-transitory computer-readable recording medium stores a bitstream of a video which is generated by a method performed by an apparatus for video processing. The method comprises: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and generating, by a neural network (NN)-based model, 40 F1254298PCTthe bitstream based on the first latent representation and the second latent representation, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0069] According to still further embodiments of the present disclosure, a method for storing bitstream of a video is provided. The method comprises: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; generating, by a neural network (NN)-based model, the bitstream based on the first latent representation and the second latent representation; and storing the bitstream in a non-transitory computer-readable recording medium, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0070] Implementations of the present disclosure can be described in view of the following clauses, the features of which can be combined in any reasonable manner.

[0071] Clause 1. A method for video processing, comprising: determining, for a conversion between a current video unit of a video and a bitstream of the video with a neural network (NN)-based model comprising a first module and a second module, a first latent representation associated with motion information of the current video unit, by the first module with a first plurality of parameters; determining, by the second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and performing the conversion based on the first latent representation and the second latent representation, wherein the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0072] Clause 2. The method of clause 1, wherein, the first module comprises: an autoregressive probability model configured to generate a first set of latent representations by using a first set of parameters among the first plurality of parameters; an upsampling model configured to generate the first latent representation based on the first set of latent representations by using a second set of parameters among the first plurality of parameters; and a synthesis network configured to generate the motion information based on the first latent representation by using a third set of parameters among the first plurality of parameters, 41 F1254298PCTand / or the second module comprises: an autoregressive probability model configured to generate a second set of latent representations by using a fourth set of parameters among the second plurality of parameters, an upsampling model configured to generate the second latent representation based on the second set of latent representations by using a fifth set of parameters among the second plurality of parameters, and a synthesis network configured to generate the residual information based on the second latent representation by using a sixth set of parameters among the second plurality of parameters.

[0073] Clause 3. The method of clause 2, wherein a resolution of a latent representation having a largest resolution among the first set of latent representations is lower than a resolution of a latent representation having a largest resolution among the second set of latent representations.

[0074] Clause 4. The method of clause 3, wherein the resolution of the latent representation having the largest resolution among the second set of latent representations is an integer multiple of the resolution of the latent representation having the largest resolution among the first set of latent representations.

[0075] Clause 5. The method of any of clauses 2-4, wherein a size of the autoregressive probability model of the first module is smaller than a size of the autoregressive probability model of the second module, and / or a size of the synthesis network of the first module is smaller than a size of the synthesis network of the second module.

[0076] Clause 6. The method of any of clauses 2-5, wherein the first module further comprises a bilinear upsampling model configured to upsample an output of the synthesis network of the first module, and a resolution of an output of the bilinear upsampling is same as a resolution of an output of the synthesis network of the second module.

[0077] Clause 7. The method of any of clauses 2-6, wherein the motion information comprises respective optical flows between the current video unit and a plurality of reference video units of the current video unit, and performing the conversion comprises: generating the respective optical flows based on the first latent representation; determining a plurality of predictions for the current video unit based on the plurality of reference video units and the respective optical flows; and determining a target prediction for the current video unit based on the plurality of predictions.

[0078] Clause 8. The method of clause 7, wherein the target prediction is determined based on a weighted sum of the plurality of predictions, and a plurality of weight maps for determining the weighted sum is dependent on predicted signals within the plurality of predictions. 42 F1254298PCT

[0079] Clause 9. The method of clause 8, wherein at least a part of the plurality of weight maps is generated by the first module or the second module.

[0080] Clause 10. The method of clause 9, wherein (N-1) weight maps among the plurality of weight maps are generated by the first module or the second module, N represents a number of the plurality of weight maps, and a remaining one of the plurality of weight maps is determined based on the (N-1) weight maps.

[0081] Clause 11. The method of any of clauses 2-10, wherein in a training process, at least one of the following is determined by training the first module and the second module based on a rate-distortion (RD) cost metric for the current video unit: the first set of latent representations, the first set of parameters, the second set of parameters, the third set of parameters, the second set of latent representations, the fourth set of parameters, the fifth set of parameters, or the sixth set of parameters.

[0082] Clause 12. The method of clause 11, wherein the RD cost metric comprises at least one of the following: an optical flow distortion term indicating a distortion between at least one optical flow generated by the first module and a ground truth for the at least one optical flow, the at least one optical flow being measured between the current video unit and at least one reference video unit of the current video unit, or a prediction distortion term indicating a distortion between at least one prediction for the current video unit and the current video unit, the at least one prediction being determined based on the at least one optical flow generated by the first module.

[0083] Clause 13. The method of clause 12, wherein the optical flow distortion term isdetermined as follows:wherein ^^^^^௪represents the optical flow distortion term, N represents a number of the at least one reference video unit,represents an optical flow between an i-th reference video unit and the current video unit that is generated by the first module,representsa ground truth for the optical flow, and ^^൫^^௧ି^→௧ , ^^ோ^ி்௧ି^→௧൯ represents a distortion between^^ and ^^ோ^ி்௧ି^→௧ ௧ି^→௧.

[0084] Clause 14. The method of any of clauses 12-13, wherein the prediction distortionterm is determined as follows:wherein ^^^^^ௗrepresents the prediction distortion term, N represents a number of the at least one reference video unit, ^^^represents a prediction for the current video unit determined based on an optical flow between an i-th reference video unit and the current video unit that 43 F1254298PCTis generated by the first module, ^^௧ represents the current video unit, and ^^(^^^ , ^^௧) representsa distortion between

[0085] Clause 15. The method of any of clauses 11-14, wherein the RD cost metric comprises at least one of the following: a first distortion term indicating a distortion associated with at least one optical flow generated by the first module, a second distortion term indicating a distortion associated with at least one prediction for the current video unit, a third distortion term indicating a distortion associated with a reconstruction for the current video unit, a first rate term indicating a rate for data associated with the first module, or a second rate term indicating a rate for data associated with the second module.

[0086] Clause 16. The method of clause 15, wherein the distortion associated with the at least one optical flow is a distortion between the at least one optical flow and a ground truth for the at least one optical flow, the at least one optical flow being measured between the current video unit and at least one reference video unit of the current video unit, and / or the distortion associated with the at least one prediction is a distortion between the at least one prediction and the current video unit, the at least one prediction being determined based on the at least one optical flow generated by the first module, and / or the distortion associated with the reconstruction is a distortion between the reconstruction and the current video unit, the reconstruction being determined based on the residual information and the at least one prediction, and / or the data associated with the first module comprise the first set of latent representations, and / or the data associated with the second module comprise the second set of latent representations.

[0087] Clause 17. The method of clause 16, wherein the first distortion term is determinedas follows:wherein ^^^௭^^௪represents the first distortion term, ^^௭[^^^^represents an optical flow betweenan i-th reference video unit and the current video unit that is generated by the first module,^^ோ^ி் represen [ ^ ோ^ி்^ ts a ground truth for the optical flow, ^^൫^^௭ ^^^ , ^^^ ൯ represents a distortionbetween ^^௭[^^^^ and ^^^ோ^ி், ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit.

[0088] Clause 18. The method of any of clauses 16-17, wherein the second distortion termis determined as follows:^^௭^^^ௗ = ∑^^(^^^^^^^^(^^^௧ି^ ,^^௭[^^^^), ^^௧),wherein ^^^௭^^ௗrepresents the second distortion term, ^^^௧ି^represents a reconstruction for an i-th reference video unit, ^^௭[^^^^ represents an optical flow between an i-th reference video 44 F1254298PCTunit and the current video unit that is generated by the first module, ^^^^^^^^(^^^௧ି^ , ^^௭[^^^^)represents an i-th prediction for the current video unit obtained by warping ^^^௧ି^ using ^^௭[^^^^,^^௧ represents the current video unit, ^^(^^, ^^) represents a distortion between A and B, and ∑represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit.

[0089] Clause 19. The method of any of clauses 16-18, wherein the third distortion termis determined as follows:^^^,௭^^^ = ^^(^^^[^^௧^ + ^^^[^^^ ⊙ ∑^^^[^^^^ ⊙ ^^^^^^^^(^^^௧ି^ , ^^௭[^^^^), ^^௧),wherein ^^^,௭^^^represents the third distortion term, ^^^[^^௧^represents the residual information generated by the second module, ^^^[^^^represents an intra-inter weight map generated by thesecond module, ⊙ represents an element-wise multiplication, ^^^௧ି^ represents areconstruction for an i-th reference video unit, ^^௭[^^^^represents an optical flow between ani-th reference video unit and the current video unit that is generated by the first module,^^^^^^^^(^^^௧ି^ ,^^௭[^^^^)represents an i-th prediction for the current video unit obtained by warping^^^௧ି^ using ^^௭[^^^^,represents a weight map for the i-th prediction generated by thesecond module, ^^௧represents the current video unit, ^^(^^,^^)represents a distortion between A and B, and ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit.

[0090] Clause 20. The method of any of clauses 16-19, wherein the first rate term isdetermined as follows:wherein ^^^௭^௧^^^represents the first rate term, ^^^௧௭represents the first set of latent representations, and௭represents a distribution of ^^^௧.

[0091] Clause 21. The method of any of clauses 16-20, wherein the second rate term isdetermined as follows:wherein ^^^^^^^ௗ௨^represents the second rate term, ^^^௧^represents the second set of latent representations, and ^^ట^(^^^௧^) represents a distribution of ^^^௧^.

[0092] Clause 22. The method of any of clauses 15-21, wherein different RD cost metrics are applied for different pictures of the video, and / or different RD cost metrics are applied for different training stages of the training process.

[0093] Clause 23. The method of any of clauses 15-22, wherein the current video unit is comprised in a t-th picture of the video, and the RD cost metric for the current video unit is defined as follows: 45 F1254298PCT^wherein ^^^^^^ represents the RD cost metric, ^^^௭^^௪represents the first distortion term, ^^^௭^^ௗrepresents the second distortion term, ^^^,௭^^^represents the third distortion term, ^^^௭^௧^^^represents the first rate term, ^^^^^^^ௗ௨^represents the second rate term, ^^^^^௪represents a weight for ^^^௭^^௪, ^^^^^ௗrepresents a weight for ^^^௭^^ௗ,represents a parameter govern a rate-distortion trade-off for the t-th picture.

[0094] Clause 24. The method of clause 23, wherein the weights ^^^^^௪and ^^^^^ௗare determined as follows: 0.1, 0 if ^^^^^^ ∈ [0, ^^^^^௪)^^^^^௪,^^^^^ௗ = ^0, 0.1 if ^^^^^^ ∈ [^^^^^௪, ^^^^^௪ + ^^^^^ௗ)0, 0 if ^^^^^^ ∈ [^^^^^௪ + ^^^^^ௗ , ^^௧^௧^^)wherein ^^^^^^ represents a current training iteration number during the training process, ^^^^^௪represents a number of training iterations in which a flow constraint is imposed, ^^^^^ௗrepresents a number of training iterations in which a prediction constraint is imposed, and^^௧^௧^^ represents a total number of training iterations performed in the training process.

[0095] Clause 25. The method of any of clauses 23-24, wherein the parameter ^^௧isdetermined as follows:wherein ^^^^^^represents a base value, ^^ represents an adjusting factor unequal to 1, ^^^^^^^^ represents a positive integer, and % represents a modulo operation.

[0096] Clause 26. The method of any of clauses 12-25, wherein the ground truth for the at least one optical flow is generated by a pretrained model.

[0097] Clause 27. The method of clause 26, wherein the pretrained model is based on recurrent all-pairs field transforms (RAFT).

[0098] Clause 28. The method of any of clauses 1-27, wherein the current video unit is a picture of the video or a portion of the picture.

[0099] Clause 29. The method of any of clauses 1-28, wherein the conversion includes decoding the current video unit from the bitstream.

[0100] Clause 30. An apparatus for video processing comprising a processor and a non- transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform a method in accordance with any of clauses 1-29.

[0101] Clause 31. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of clauses 1-29. 46 F1254298PCT

[0102] Clause 32. A non-transitory computer-readable recording medium storing a bitstream of a video which is generated by a method performed by an apparatus for video processing, wherein the method comprises: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and generating, by a neural network (NN)-based model, the bitstream based on the first latent representation and the second latent representation, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

[0103] Clause 33. A method for storing a bitstream of a video, comprising: determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; generating, by a neural network (NN)-based model, the bitstream based on the first latent representation and the second latent representation; and storing the bitstream in a non-transitory computer-readable recording medium, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream. Example Device

[0104] Fig. 6 illustrates a block diagram of a computing device 600 in which various embodiments of the present disclosure can be implemented. The computing device 600 may be implemented as or included in the source device 110 (or the visual data encoder 114) or the destination device 120 (or the visual data decoder 124).

[0105] It would be appreciated that the computing device 600 shown in Fig. 6 is merely for purpose of illustration, without suggesting any limitation to the functions and scopes of the embodiments of the present disclosure in any manner.

[0106] As shown in Fig. 6, the computing device 600 includes a general-purpose computing device 600. The computing device 600 may at least comprise one or more processors or processing units 610, a memory 620, a storage unit 630, one or more communication units 640, one or more input devices 650, and one or more output devices 660. 47 F1254298PCT

[0107] In some embodiments, the computing device 600 may be implemented as any user terminal or server terminal having the computing capability. The server terminal may be a server, a large-scale computing device or the like that is provided by a service provider. The user terminal may for example be any type of mobile terminal, fixed terminal, or portable terminal, including a mobile phone, station, unit, device, multimedia computer, multimedia tablet, Internet node, communicator, desktop computer, laptop computer, notebook computer, netbook computer, tablet computer, personal communication system (PCS) device, personal navigation device, personal digital assistant (PDA), audio / video player, digital camera / video camera, positioning device, television receiver, radio broadcast receiver, E-book device, gaming device, or any combination thereof, including the accessories and peripherals of these devices, or any combination thereof. It would be contemplated that the computing device 600 can support any type of interface to a user (such as “wearable” circuitry and the like).

[0108] The processing unit 610 may be a physical or virtual processor and can implement various processes based on programs stored in the memory 620. In a multi-processor system, multiple processing units execute computer executable instructions in parallel so as to improve the parallel processing capability of the computing device 600. The processing unit 610 may also be referred to as a central processing unit (CPU), a microprocessor, a controller or a microcontroller.

[0109] The computing device 600 typically includes various computer storage medium. Such medium can be any medium accessible by the computing device 600, including, but not limited to, volatile and non-volatile medium, or detachable and non-detachable medium. The memory 620 can be a volatile memory (for example, a register, cache, Random Access Memory (RAM)), a non-volatile memory (such as a Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), or a flash memory), or any combination thereof. The storage unit 630 may be any detachable or non-detachable medium and may include a machine-readable medium such as a memory, flash memory drive, magnetic disk or another other media, which can be used for storing information and / or visual data and can be accessed in the computing device 600.

[0110] The computing device 600 may further include additional detachable / non- detachable, volatile / non-volatile memory medium. Although not shown in Fig. 6, it is possible to provide a magnetic disk drive for reading from and / or writing into a detachable and non-volatile magnetic disk and an optical disk drive for reading from and / or writing into a detachable non-volatile optical disk. In such cases, each drive may be connected to a bus 48 F1254298PCT(not shown) via one or more visual data medium interfaces.

[0111] The communication unit 640 communicates with a further computing device via the communication medium. In addition, the functions of the components in the computing device 600 can be implemented by a single computing cluster or multiple computing machines that can communicate via communication connections. Therefore, the computing device 600 can operate in a networked environment using a logical connection with one or more other servers, networked personal computers (PCs) or further general network nodes.

[0112] The input device 650 may be one or more of a variety of input devices, such as a mouse, keyboard, tracking ball, voice-input device, and the like. The output device 660 may be one or more of a variety of output devices, such as a display, loudspeaker, printer, and the like. By means of the communication unit 640, the computing device 600 can further communicate with one or more external devices (not shown) such as the storage devices and display device, with one or more devices enabling the user to interact with the computing device 600, or any devices (such as a network card, a modem and the like) enabling the computing device 600 to communicate with one or more other computing devices, if required. Such communication can be performed via input / output (I / O) interfaces (not shown).

[0113] In some embodiments, instead of being integrated in a single device, some or all components of the computing device 600 may also be arranged in cloud computing architecture. In the cloud computing architecture, the components may be provided remotely and work together to implement the functionalities described in the present disclosure. In some embodiments, cloud computing provides computing, software, visual data access and storage service, which will not require end users to be aware of the physical locations or configurations of the systems or hardware providing these services. In various embodiments, the cloud computing provides the services via a wide area network (such as Internet) using suitable protocols. For example, a cloud computing provider provides applications over the wide area network, which can be accessed through a web browser or any other computing components. The software or components of the cloud computing architecture and corresponding visual data may be stored on a server at a remote position. The computing resources in the cloud computing environment may be merged or distributed at locations in a remote visual data center. Cloud computing infrastructures may provide the services through a shared visual data center, though they behave as a single access point for the users. Therefore, the cloud computing architectures may be used to provide the components and functionalities described herein from a service provider at a remote location. Alternatively, they may be provided from a conventional server or installed directly or otherwise on a client 49 F1254298PCTdevice.

[0114] The computing device 600 may be used to implement visual data encoding / decoding in embodiments of the present disclosure. The memory 620 may include one or more visual data coding modules 625 having one or more program instructions. These modules are accessible and executable by the processing unit 610 to perform the functionalities of the various embodiments described herein.

[0115] In the example embodiments of performing visual data encoding, the input device 650 may receive visual data as an input 670 to be encoded. The visual data may be processed, for example, by the visual data coding module 625, to generate an encoded bitstream. The encoded bitstream may be provided via the output device 660 as an output 680.

[0116] In the example embodiments of performing visual data decoding, the input device 650 may receive an encoded bitstream as the input 670. The encoded bitstream may be processed, for example, by the visual data coding module 625, to generate decoded visual data. The decoded visual data may be provided via the output device 660 as the output 680.

[0117] While this disclosure has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present application as defined by the appended claims. Such variations are intended to be covered by the scope of this present application. As such, the foregoing description of embodiments of the present application is not intended to be limiting. 50 F1254298PCT

Claims

1. I / We Claim:

1. A method for video processing, comprising: determining, for a conversion between a current video unit of a video and a bitstream ofthe video with a neural network (NN)-based model comprising a first module and a second module,a first latent representation associated with motion information of the current video unit, by the first module with a first plurality of parameters; determining, by the second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and performing the conversion based on the first latent representation and the second latentrepresentation, wherein the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

2. The method of claim 1, wherein,the first module comprises: an autoregressive probability model configured to generate a first set of latent representations by using a first set of parameters among the first plurality of parameters; an upsampling model configured to generate the first latent representation based on the first set of latent representations by using a second set of parameters among the firstplurality of parameters; and a synthesis network configured to generate the motion information based on the first latent representation by using a third set of parameters among the first plurality of parameters, and / or the second module comprises:an autoregressive probability model configured to generate a second set of latent representations by using a fourth set of parameters among the second plurality of parameters, an upsampling model configured to generate the second latent representation based on the second set of latent representations by using a fifth set of parameters among thesecond plurality of parameters, and a synthesis network configured to generate the residual information based on the second latent representation by using a sixth set of parameters among the second plurality of parameters. 51 F1254298PCT3. The method of claim 2, wherein a resolution of a latent representation having a largest resolution among the first set of latent representations is lower than a resolution of a latent representation having a largest resolution among the second set of latent representations.

4. The method of claim 3, wherein the resolution of the latent representation having the largest resolution among the second set of latent representations is an integer multiple of the resolution of the latent representation having the largest resolution among the first set of latent representations.

5. The method of any of claims 2-4, wherein a size of the autoregressive probability model of the first module is smaller than a size of the autoregressive probability model of the second module, and / or a size of the synthesis network of the first module is smaller than a size of the synthesisnetwork of the second module.

6. The method of any of claims 2-5, wherein the first module further comprises a bilinear upsampling model configured to upsample an output of the synthesis network of the first module, and a resolution of an output of the bilinear upsampling is same as a resolution of an output of thesynthesis network of the second module.

7. The method of any of claims 2-6, wherein the motion information comprises respective optical flows between the current video unit and a plurality of reference video units of the current video unit, and performing the conversion comprises: generating the respective optical flows based on the first latent representation;determining a plurality of predictions for the current video unit based on the plurality of reference video units and the respective optical flows; and determining a target prediction for the current video unit based on the plurality of predictions.

8. The method of claim 7, wherein the target prediction is determined based on a weighted sum of the plurality of predictions, and a plurality of weight maps for determining the weighted sum is dependent on predicted signals within the plurality of predictions. 52 F1254298PCT9. The method of claim 8, wherein at least a part of the plurality of weight maps is generated by the first module or the second module.

10. The method of claim 9, wherein (N-1) weight maps among the plurality of weight mapsare generated by the first module or the second module, N represents a number of the plurality ofweight maps, and a remaining one of the plurality of weight maps is determined based on the (N- 1) weight maps.

11. The method of any of claims 2-10, wherein in a training process, at least one of thefollowing is determined by training the first module and the second module based on a rate-distortion (RD) cost metric for the current video unit: the first set of latent representations, the first set of parameters, the second set of parameters, the third set of parameters,the second set of latent representations, the fourth set of parameters, the fifth set of parameters, or the sixth set of parameters.

12. The method of claim 11, wherein the RD cost metric comprises at least one of the following: an optical flow distortion term indicating a distortion between at least one optical flow generated by the first module and a ground truth for the at least one optical flow, the at least oneoptical flow being measured between the current video unit and at least one reference video unitof the current video unit, or a prediction distortion term indicating a distortion between at least one prediction for the current video unit and the current video unit, the at least one prediction being determined based on the at least one optical flow generated by the first module.

13. The method of claim 12, wherein the optical flow distortion term is determined asfollows:wherein ^^^^^௪represents the optical flow distortion term, N represents a number of the at 53 F1254298PCTleast one reference video unit, ^^௧ି^→௧represents an optical flow between an i-th reference video unit and the current video unit that is generated by the first module,represents a ground truth for the optical flow, andrepresents a distortion between14. The method of any of claims 12-13, wherein the prediction distortion term isdetermined as follows:^^^^^ௗ = ∑^ୀே^ୀ^ ^^ (^^^ , ^^௧),wherein ^^^^^ௗrepresents the prediction distortion term, N represents a number of the at least one reference video unit,^^^represents a prediction for the current video unit determinedbased on an optical flow between an i-th reference video unit and the current video unit that isgenerated by the first module, ^^௧ represents the current video unit, and ^^(^^^, ^^௧) represents adistortion between ^^^and ^^௧.

15. The method of any of claims 11-14, wherein the RD cost metric comprises at least oneof the following:a first distortion term indicating a distortion associated with at least one optical flow generated by the first module, a second distortion term indicating a distortion associated with at least one prediction for the current video unit, athird distortion term indicating a distortion associated with a reconstruction for the currentvideo unit, a first rate term indicating a rate for data associated with the first module, or a second rate term indicating a rate for data associated with the second module.

16. The method of claim 15, wherein the distortion associated with the at least one opticalflow is a distortion between the at least one optical flow and a ground truth for the at least one optical flow, the at least one optical flow being measured between the current video unit and at least one reference video unit of the current video unit, and / or the distortion associated with the at least one prediction is a distortion between the at leastone prediction and the current video unit, the at least one prediction being determined based onthe at least one optical flow generated by the first module, and / or the distortion associated with the reconstruction is a distortion between the reconstruction and the current video unit, the reconstruction being determined based on the residual information and the at least one prediction, and / or 54 F1254298PCTthe data associated with the first module comprise the first set of latent representations, and / or the data associated with the second module comprise the second set of latent representations.

17. The method of claim 16, wherein the first distortion term is determined as follows:wherein ^^௭^^^௪ represents the first distortion term, ^^௭[^^^^ represents an optical flowbetween an i-th reference video unit and the current video unit that is generated by the first module,^^ோ^ி்^ represents a ground truth for the optical flow, ^^൫^^௭[^^^^, ^^ோ^ி்^ ൯ represents a distortionbetween ^^௭[^^^^ and^^^ோ^ி், ∑ represents a summation for i from 1 to N, and N represents a number of the at least one reference video unit.

18. The method of any of claims 16-17, wherein the second distortion term is determinedas follows:^^௭^^^ௗ = ∑^^(^^^^^^^^(^^^௧ି^ , ^^௭[^^^^), ^^௧),wherein^^^௭^^ௗrepresents the second distortion term, ^^^௧ି^represents a reconstruction for an i-th reference video unit, ^^௭[^^^^ represents an optical flow between an i-th reference video unitand the current video unit that is generated by the first module,^^^^^^^^(^^^௧ି^ , ^^௭[^^^^) represents ani-th prediction for the current video unit obtained by warping ^^^௧ି^ using ^^௭[^^^^, ^^௧ represents thecurrent video unit, ^^(^^,^^) represents a distortion between A and B, and ∑ represents asummation for i from 1 to N, and N represents a number of the at least one reference video unit.

19. The method of any of claims 16-18, wherein the third distortion term is determined asfollows:^^^,௭^^^ = ^^(^^^[^^௧^ + ^^^[^^^ ⊙ ∑^^^[^^^^ ⊙ ^^^^^^^^(^^^௧ି^ ,^^௭[^^^^), ^^௧),wherein ^^^,௭^^^ represents the third distortion term, ^^^[^^௧^ represents the residualinformation generated by the second module,^^^[^^^ represents an intra-inter weight map generatedby the second module, ⊙ represents an element-wise multiplication, ^^^௧ି^ represents areconstruction for an i-th reference video unit, ^^௭[^^^^ represents an optical flow between an i-threference video unit and the current video unit that is generated by the first module,^^^^^^^^(^^^௧ି^,^^௭[^^^^) represents an i-th prediction for the current video unit obtained by warping^^^௧ି^ using ^^௭[^^^^, ^^^represents a weight map for the i-th prediction generated by the second55 F1254298PCTmodule, ^^௧ represents the current video unit, ^^(^^, ^^) represents a distortion between A and B, and∑ represents a summation for i from 1 to N, and N represents a number of the at least one referencevideo unit.

20. The method of any of claims 16-19, wherein the first rate term is determined as follows:wherein ^^^௭^௧^^^represents the first rate term, ^^^௧௭represents the first set of latentrepresentations, and ^^ట^(^^^௧௭)represents a distribution of ^^^௧௭.

21. The method of any of claims 16-20, wherein the second rate term is determined asfollows:^^^^^^^ௗ௨^ = −^^^^^^ଶ^^ట^(^^^^௧ ),wherein ^^^^^^^ௗ௨^represents the second rate term, ^^^௧^represents the second set of latentrepresentations, and^^ట^(^^^௧^)represents a distribution of ^^^௧^.

22. The method of any of claims 15-21, wherein different RD cost metrics are applied for different pictures of the video, and / or different RD cost metrics are applied for different training stages of the training process.

23. The method of any of claims 15-22, wherein the current video unit is comprised in a t-th picture of the video, and the RD cost metric for the current video unit is defined as follows:wherein ^^^^^^ represents the RD cost metric, ^^௭^^^௪ represents the first distortion term,^^௭ represents the seco ^,௭ ௭^^^ௗ nd distortion term, ^^^^^ represents the third distortion term, ^^^^௧^^^represents the first rate term, ^^^^^^^ௗ௨^ represents the second rate term,^^^^^௪ represents a weightfor^^^௭^^௪ ,^^^^^ௗ represents a weight for^^^௭^^ௗ ,represents a parameter govern a rate-distortion trade-off for the t-th picture.

24. The method of claim 23, wherein the weights^^^^^௪and^^^^^ௗare determined asfollows:56 F1254298PCTwherein^^^^^^represents a current training iteration number during the training process, ^^^^^௪represents a number of training iterations in which a flow constraint is imposed,^^^^^ௗrepresents a number of training iterations in which a prediction constraint is imposed, and ^^௧^௧^^represents a total number of training iterations performed in the training process.

25. The method of any of claims 23-24, wherein the parameter ^^௧ is determined as follows:wherein ^^^^^^represents a base value, ^^ represents an adjusting factor unequal to 1, ^^^^^^^^ represents a positive integer, and%represents a modulo operation.

26. The method of any of claims 12-25, wherein the ground truth for the at least one optical flow is generated by a pretrained model.

27. The method of claim 26, wherein the pretrained model is based on recurrent all-pairsfield transforms (RAFT).

28. The method of any of claims 1-27, wherein the current video unit is a picture of the video or a portion of the picture.

29. The method of any of claims 1-28, wherein the conversion includes decoding thecurrent video unit from the bitstream.

30. An apparatus for video processing comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause theprocessor to perform a method in accordance with any of claims 1-29.

31. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of claims 1-29.

32. A non-transitory computer-readable recording medium storing a bitstream of a videowhich is generated by a method performed by an apparatus for video processing, wherein the method comprises: determining, by a first module with a first plurality of parameters, a first latent 57 F1254298PCTrepresentation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; and generating, by a neural network (NN)-based model, the bitstream based on the first latentrepresentation and the second latent representation,wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and the second plurality of parameters are indicated in the bitstream.

33. A method for storing a bitstream of a video, comprising:determining, by a first module with a first plurality of parameters, a first latent representation associated with motion information of a current video unit of the video; determining, by a second module with a second plurality of parameters, a second latent representation associated with residual information of the current video unit; generating, by a neural network (NN)-based model, the bitstream based on the first latentrepresentation and the second latent representation; and storing the bitstream in a non-transitory computer-readable recording medium, wherein the first module and the second module are comprised in the NN-based model, the second module is independent from the first module, and the first plurality of parameters and thesecond plurality of parameters are indicated in the bitstream.58 F1254298PCT

Images