Efficient processing, i.e., analysis, storage, access and transmission of visual content, with continuously increasing data rates, in environments which are more and more mobile and distributed, remains a key challenge of the signal and image processing community. New imaging modalities, High Dynamic Range (HDR) imaging, multiview, plenoptic, light fields, 360o videos, generating very large volumes of data contribute to the sustained need for efficient algorithms for a variety of processing tasks.

Building upon a strong background on signal/image/video processing and information theory, the goal of the SIROCCO team is to design mathematically founded tools and algorithms for visual data analysis, modeling, representation, coding, and processing, with for the latter area an emphasis on inverse problems related to super-resolution, view synthesis, HDR recovery from multiple exposures, denoising and inpainting. Even if 2D imaging is still within our scope, the goal is to give a particular attention to HDR imaging, light fields, and 360o videos.
The project-team activities are structured and organized around the following inter-dependent research axes:

While aiming at generic approaches, some of the solutions developed are applied to practical problems in partnership with industry (InterDigital, Ateme, Orange) or in the framework of national projects. The application domains addressed by the project are networked visual applications taking into account their various requirements and needs in terms of compression, of network adaptation, of advanced functionalities such as navigation, interactive streaming and high quality rendering.

Most visual data processing problems require a prior step of data analysis, of discovery and modeling of correlation structures. This is a pre-requisite for the design of dimensionality reduction methods, of compact representations and of fast processing techniques. These correlation structures often depend on the scene and on the acquisition system. Scene analysis and modeling from the data at hand is hence also part of our activities. To give examples, scene depth and scene flow estimation is a cornerstone of many approaches in multi-view and light field processing. The information on scene geometry helps constructing representations of reduced dimension for efficient (e.g. in interactive time) processing of new imaging modalities (e.g. light fields or 360o videos).

Dimensionality reduction has been at the core of signal and image processing methods, for a number of years now, hence have obviously always been central to the research of Sirocco. These methods encompass sparse and low-rank models, random low-dimensional projections in a compressive sensing framework, and graphs as a way of representing data dependencies and defining the support for learning and applying signal de-correlating transforms. The study of these models and signal processing tools is even more compelling for designing efficient algorithms for processing the large volumes of high-dimensionality data produced by novel imaging modalities. The models need to be adapted to the data at hand through learning of dictionaries or of neural networks. In order to define and learn local low-dimensional or sparse models, it is necessay to capture and understand the underlying data geometry, e.g. with the help of manifolds and manifold clustering tools. It also requires exploiting the scene geometry with the help of disparity or depth maps, or its variations in time via coarse or dense scene flows.

Based on the above models, besides compression, our goal is also to develop algorithms for solving a number of inverse problems in computer vision. Our emphasis is on methods to cope with limitations of sensors (e.g. enhancing spatial, angular or temporal resolution of captured data, or noise removal), to synthesize virtual views or to reconstruct (e.g. in a compressive sensing framework) light fields from a sparse set of input views, to recover HDR visual content from multiple exposures, and to enable content editing (we focus on color transfer, re-colorization, object removal and inpainting). Note that view synthesis is a key component of multiview and light field compression. View synthesis is also needed to support user navigation and interactive streaming. It is also needed to avoid angular aliasing in some post-capture processing tasks, such as re-focusing, from a sparse light field. Learning models for the data at hand is key for solving the above problems.

The ever-growing volume of image/video traffic motivates the search for new coding solutions suitable for band and energy limited networks but also space and energy limited storage devices. In particular, we investigate compression strategies that are adapted to the users needs and data access requests in order to meet all these transmission and/or storage constraints. Our first goal is to address theoretical issues such as the information theoretical bounds of these compression problems. This includes compression of a database with random access, compression with interactivity, and also data repurposing that takes into account the users needs and user data perception. A second goal is to construct practical coding for all these problems.

The research activities on analysis, compression and communication of visual data mostly rely on tools and formalisms from the areas of statistical image modeling, of signal processing, of machine learning, of coding and information theory. Some of the proposed research axes are also based on scientific foundations of computer vision (e.g. multi-view modeling and coding). We have limited this section to some tools which are central to the proposed research axes, but the design of complete compression and communication solutions obviously rely on a large number of other results in the areas of motion analysis, transform design, entropy code design, etc which cannot be all described here.

Dimensionality reduction encompasses a variety of methods for low-dimensional data embedding, such as sparse and low-rank models, random low-dimensional projections in a compressive sensing framework, and sparsifying transforms including graph-based transforms. These methods are the cornerstones of many visual data processing tasks (compression, inverse problems).

Sparse representations, compressive sensing, and dictionary learning have been shown to be powerful tools for efficient processing of visual data. The objective of sparse representations is to find a sparse approximation of a given input data. In theory, given a dictionary matrix

The recent theory of compressed sensing, in the context of discrete signals, can be seen as an effective dimensionality reduction technique.
The idea behind compressive sensing is that
a signal can be accurately recovered from a small number of linear measurements, at a rate much smaller than what is commonly prescribed by the Shannon-Nyquist theorem, provided that it is sparse or compressible in a known basis. Compressed sensing has emerged as a powerful framework for signal acquisition and sensor design, with a number of open issues such as learning the basis in which the signal is sparse, with the help of dictionary learning methods, or the design and optimization of the sensing matrix. The problem is in particular investigated in the context of light fields acquisition, aiming at novel camera design with the goal of offering a good trade-off between spatial and angular resolution.

While most image and video processing methods have been developed for cartesian sampling grids, new imaging modalities (e.g. point clouds, light fields) call for representations on irregular supports that can be well represented by graphs. Reducing the dimensionality of such signals require designing novel transforms yielding compact signal representation.
One example of transform is the Graph Fourier transform
whose basis functions are given by the eigenvectors of the graph Laplacian matrix

From dictionary learning which we have investigated a lot in the past, our activity is now evolving towards deep learning techniques which we are considering for dimensionality reduction. We address the problem of unsupervised learning of transforms and prediction operators that would be optimal in terms of energy compaction, considering autoencoders and neural network architectures.

An autoencoder is a neural network with an encoder

To avoid this limitation, architectures without fully-connected layer and comprising instead convolutional layers and non-linear operators, forming convolutional neural networks (CNN) may be preferrable. The obtained representation is thus a set of so-called feature maps.

The other problems that we address with the help of neural networks are scene geometry and scene flow estimation, view synthesis, prediction and interpolation with various imaging modalities. The problems are posed either as supervised or unsupervised learning tasks. Our scope of investigation includes autoencoders, convolutional networks, variational autoencoders and generative adversarial networks (GAN) but also recurrent networks and in particular Long Short Term Memory (LSTM) networks. Recurrent neural networks attempting to model time or sequence dependent behaviour, by feeding back the output of a neural network layer at time t to the input of the same network layer at time t+1, have been shown to be interesting tools for temporal frame prediction. LSTMs are particular cases of recurrent networks made of cells composed of three types of neural layers called gates.

Deep neural networks have also been shown to be very promising for solving inverse problems (e.g. super-resolution, sparse recovery in a compressive sensing framework, inpainting) in image processing. Variational autoencoders, generative adversarial networks (GAN), learn, from a set of examples, the latent space or the manifold in which the images, that we search to recover, reside. The inverse problems can be re-formulated using a regularization in the latent space learned by the network. For the needs of the regularization, the learned latent space may need to verify certain properties such as preserving distances or neighborhood of the input space, or in terms of statistical modeling. GANs, trained to produce images that are plausible, are also useful tools for learning texture models, expressed via the filters of the network, that can be used for solving problems like inpainting or view synthesis.

Source coding and channel coding theory 1 is central to our compression and communication activities, in particular to the design of entropy codes and of error correcting codes. Another field in coding theory which has emerged in the context of sensor networks is Distributed Source Coding (DSC). It refers to the compression of correlated signals captured by different sensors which do not communicate between themselves. All the signals captured are compressed independently and transmitted to a central base station which has the capability to decode them jointly. DSC finds its foundation in the seminal Slepian-Wolf 2 (SW) and Wyner-Ziv 3 (WZ) theorems. Let us consider two binary correlated sources

In 1976, Wyner and Ziv considered the problem of coding of two correlated sources

The application domains addressed by the project are:

Compression of visual content remains a widely-sought capability for a large number of applications. This is particularly true for mobile applications, as the need for wireless transmission capacity will significantly increase during the years to come. Hence, efficient compression tools are required to satisfy the trend towards mobile access to larger image resolutions and higher quality. A new impulse to research in video compression is also brought by the emergence of new imaging modalities, e.g. high dynamic range (HDR) images and videos (higher bit depth, extended colorimetric space), light fields and omni-directional imaging.

Different video data formats and technologies are envisaged for interactive and immersive 3D video applications using omni-directional videos, stereoscopic or multi-view videos. The "omni-directional video" set-up refers to 360-degree view from one single viewpoint or spherical video. Stereoscopic video is composed of two-view videos, the right and left images of the scene which, when combined, can recreate the depth aspect of the scene. A multi-view video refers to multiple video sequences captured by multiple video cameras and possibly by depth cameras. Associated with a view synthesis method, a multi-view video allows the generation of virtual views of the scene from any viewpoint. This property can be used in a large diversity of applications, including Three-Dimensional TV (3DTV), and Free Viewpoint Video (FVV). In parallel, the advent of a variety of heterogeneous delivery infrastructures has given momentum to extensive work on optimizing the end-to-end delivery QoS (Quality of Service). This encompasses compression capability but also capability for adapting the compressed streams to varying network conditions. The scalability of the video content compressed representation and its robustness to transmission impairments are thus important features for seamless adaptation to varying network conditions and to terminal capabilities.

Free-viewpoint Television (FTV) is a system for watching videos in which the user can choose its viewpoint freely and change it at anytime. To allow this navigation, many views are proposed and the user can navigate from one to the other. The goal of FTV is to propose an immersive sensation without the disadvantage of Three-dimensional television (3DTV). With FTV, a look-around effect is produced without any visual fatigue since the displayed images remain 2D. However, technical characteristics of FTV are large databases, huge numbers of users, and requests of subsets of the data, while the subset can be randomly chosen by the viewer. This requires the design of coding algorithms allowing such a random access to the pre-encoded and stored data which would preserve the compression performance of predictive coding. This research also finds applications in the context of Internet of Things in which the problem arises of optimally selecting both the number and the position of reference sensors and of compressing the captured data to be shared among a high number of users.

Broadband fixed and mobile access networks with different radio access technologies have enabled not only IPTV and Internet TV but also the emergence of mobile TV and mobile devices with internet capability. A major challenge for next internet TV or internet video remains to be able to deliver the increasing variety of media (including more and more bandwidth demanding media) with a sufficient end-to-end QoS (Quality of Service) and QoE (Quality of Experience).

Editing and post-production are critical aspects in the audio-visual production process. Increased ways of “consuming” visual content also highlight the need for content repurposing as well as for higher interaction and editing capabilities. Content repurposing encompasses format conversion (retargeting), content summarization, and content editing. This processing requires powerful methods for extracting condensed video representations as well as powerful inpainting techniques. By providing advanced models, advanced video processing and image analysis tools, more visual effects, with more realism become possible. Our activies around light field imaging also find applications in computational photography which refers to the capability of creating photographic functionalities beyond what is possible with traditional cameras and processing tools.

No social or environmental responsibility.

This year has seen the start of

This section describes the new software developed in the year 2021 as well as the datasets created and the platform under development.

This code implements a new image compression algorithm that allows to navigate within a static scene. To do so, the code provides access in the compressed domain to any block and therefore allows extraction of any subpart of the image. This codec implements this interactive compression for two image modalities: omnidirectional images and texture maps of 3D models. For omnidirectional images the input is a 2D equirectangular projection of the 360 image. The output is the image seen in the viewport. For 3D models, the input is a texture map and the 3D mesh. The output is also the image seen in the viewport.

The code consists of three parts: (A) an offline encoder (B) an online bit extractor and (C) a decoder. The offline encoder (i) partitions the image into blocks, (ii) optimizes the positions of the access blocks, (iii) computes a set of geometry aware predictions for each block (to cover all possible navigation paths), (iv) implements transform quantization for all blocks and their predictions, and finally (v) evaluates the encoding rates. The online bit extractor (Part B) first computes the optimal and geometry aware scanning order. Then it extracts in the bitstream, the sufficient amount of information to allow the decoding of the requested blocks. The last part of the code is the decoder (Part C). The decoder reconstructs the same scanning order as the one computed at the online bit extractor. Then, the blocks are decoded (inverse transform, geometry aware predictions, ...) and reconstructed. Finally the image in the viewport is generated.

This code implements a new video compression algorithm for omnidirectional (360°) videos. The main originality of this algorithm is that the compression is performed directly on the sphere. First, it saves computational complexity as it avoids to project the sphere onto a 2D map, as classically done. Second, and more importantly, it allows to achieve a better rate-distortion tradeoff, since neither the visual data nor its domain are distorted. This code implements an extension from still images to videos of the on-the-sphere compression for omnidirectional still images. In particular, it implements a novel rate-distortion optimized motion estimation algorithm to perform motion compensation. The optimization is performed among a set of existing motion models and a novel motion model called tangent-linear+t. Moreover, a finer search pattern, called spherical-uniform, is also implemented for the motion parameters, which leads to a more accurate block prediction. The novel algorithm leads to rate-distortion gains compared to methods based on a unique motion model.

More precisely, this algorithm performs inter-prediction and contains (i) a motion estimation, (ii) a motion compensation and computation of a residue, (iii) the encoding of this residue, (iv) estimation of rate and distortion. Several visualization tools are also provided such as rate-distortion curves, but also search pattern centers, and motion field.

The scientific and industrial community is nowadays exploring new multimedia applications using 3D data (beyond stereoscopy). In particular, Free Viewpoint Television (FTV) has attracted much attention in the recent years. In those systems, user can choose in real time its view angle from which he wants to observe the scene. Despite the great interest for FTV, the lack of realistic and ambitious datasets penalizes the research effort. The acquisition of such sequences is very costly in terms of hardware and working effort, which explains why no multi-view videos suitable for FTV has been proposed yet.

In the context of the project ADT ATeP 2016-2018 (funded by Inria), such datasets were acquired and some calibration tools have been developed.
First 40 omnidirectional cameras and their associated equipments have been acquired by the team (thanks to Rennes Metropole funding). We have first focused on the calibration of this camera, i.e., the development of the relationship between a 3D point and its projection in the omnidirectional image. In particular, we have shown that the unified spherical model fits the acquired omnidirectional cameras. Second, we have developed tools to calibrate the cameras in relation to each other. Finally, we have made a capture of 3 multiview sequences that have been made available to the community via a public web site.

As part of the ERC Clim project, the EPI Sirocco is developing a light field processing toolbox. The toolbox and libraries are developed in C++ and the graphical user interface relies on Qt. As input data, this tool accepts both sparse light fields acquired with High Density Camera Arrays (HDCA) and denser light fields captured with plenoptic cameras using microlens arrays (MLA). At the time of writing, in addition to some simple functionalities, such as re-focusing, change of viewpoints, with different forms of visualization, the toolbox integrates more advanced tools for scene depth estimation from sparse and dense light fields, for super-ray segmentation and scene flow estimation, and for light field denoising and angular interpolation using anisotropic diffusion in the 4D ray space. The toolbox is now being interfaced with the C/C++ API of the tensorflow platform, in order to execute deep models developed in the team for scene depth and scene flow estimation, view synthesis, and axial super-resolution.

In the Intercom project, we have studied the impact of interactivity on the coding performance. We have, for example, tackled the following problem: is it possible to compress a 360

Compressive light field photography enables light field acquisition using a single sensor by utilizing a color coded mask. This approach is very cost effective since consumer-level digital cameras can be turned into a light field camera by simply placing a coded mask between the sensor and the aperture plane and solving an inverse problem to obtain an estimate of the original light field. While in the past years, we developed solutions based on signal processing methods

8, in 2021 we have developed a deep learning architecture for compressive light field acquisition using a color coded mask and a sensor with Color Filter Array (CFA)

10, in line with the multi-mask camera model we proposed in

12. Unlike previous methods where a fixed mask pattern is used, our deep network learns the optimal distribution of the color coded mask pixels. The proposed solution enables end-to-end learning of the color-coded mask distribution and the reconstruction network, taking into account the sensor CFA. Consequently, the resulting network can efficiently perform joint demosaicing and light field reconstruction of images acquired with color-coded mask and a CFA sensor. Compared to previous methods based on deep learning with monochrome sensors, as well as traditional compressive sensing approaches using CFA sensors, we obtain superior color reconstruction of the light fields.

We have also presented an efficient and mathematically grounded deep learning model to reconstruct a light field from a set of measurements obtained using a color-coded mask and a color filter array (CFA). Following the promising trend of unrolling optimization algorithms with learned priors, we formulate our task of light field reconstruction as an inverse problem and derive a principled deep network architecture from this formulation. We also introduce a closed-form extraction of information from the acquisition, while similar methods found in the recent literature systematically use an approximation. Compared to similar deep learning methods, we show that our approach allows for a better reconstruction quality. We have further shown that our approach is robust to noise using realistic simulations of the sensing acquisition process.

The problem of HDR light field acquisition using a 2D sensor remains an open and challenging problem despite the recent advances in both 2D HDR imaging and compressive LDR light field acquisition. The main challenge here is indeed the reconstruction of an HDR light field from a single LDR image recorded on a monochrome sensor equipped with a Color Filter Array (CFA). This single monochrome image hence should encode not only the HDR information of the scene, but also angular and spectral measurements of the light field.

To address this problem, in collaboration with the Univ. of Linkoping, we have introduced a novel framework for compressive capture of HDR light fields combining multiple ISO photography with mask-based coded projection techniques 11. The approach builds upon the multi-mask camera model we proposed in 12, and based on a main lens, a multi-ISO sensor and a coded mask located in the optical path between the main lens and the sensor. The mask projects coded spatio-angular information of the light field onto the 2D sensor. Hence, our compressive HDR light field imaging framework captures a coded image with a varying per pixel gain encoding the scene. The sensor image captured through the mask, the varying per pixel gain, and the CFA, encodes spatial, angular, and color intensity variations in the scene. This coded projection image compresses the incident scene radiance information such that the full HDR light field can be recovered as a tractable inverse problem. The model encompasses different acquisition scenarios with different ISO patterns and gains. Moreover, we assume that the sensor has a built-in color filter array (CFA), making our design more suitable for consumer-level cameras.

We, in parallel, developed a novel joint spatio-angular-HDR reconstruction algorithm using a trained dictionary specifically designed for HDR light field reconstruction. The joint reconstruction includes a confidence matrix based on the pixel intensity and acquisition noise, effectively performing denoising as an integral part in the reconstruction. The reconstruction algorithm actually jointly performs color demosaicing, light field angular information recovery, HDR reconstruction, and denoising from the multi-ISO measurements formed on the sensor (see an illustration of some results in Fig.(3).

We have also created two HDR light field data sets: one synthetic data set created using the Blender rendering software with two baselines, and a real light field data set created from the fusion of multi-exposure low dynamic range (LDR) images captured using a Lytro Illum light field camera. Experimental results show that, with a sampling rate as low as 2.67 %, using two shots, our proposed method yields a higher light field reconstruction quality compared to the fusion of multiple LDR light fields captured with different exposures, and with the fusion of multiple LDR light fields captured with different ISO settings. This framework leads to a new design for single sensor compressive HDR light field cameras, combining multi-ISO photography with coded mask acquisition, placed in a compressive sensing framework.

Immersive video often refers to multiple views with texture and scene geometry information, from which different viewports can be synthesized on the client side. To design efficient immersive video coding solutions, it is desirable to minimize bitrate, pixel rate and complexity. We have investigated whether the classical approach of sending the geometry of a scene as depth maps is appropriate to serve this purpose. Previous work has shown that bypassing depth transmission entirely and estimating depth at the client side improves the synthesis performance while saving bitrate and pixel rate. In order to understand if the encoder side depth maps contain information that is beneficial to be transmitted, we have first explored a hybrid approach which enables partial depth map transmission using a block-based RD-based decision in the depth coding process 7. This approach has revealed that partial depth map transmission may improve the rendering performance but does not present a good compromise in terms of compression efficiency. This led us to address the remaining drawbacks of decoder side depth estimation: complexity and depth map inaccuracy. We propose a novel system that takes advantage of high quality depth maps at the server side by encoding them into lightweight features that support the depth estimator at the client side. These features have allowed reducing the amount of data that has to be handled during decoder side depth estimation by 88%, which significantly speeds up the cost computation and the energy minimization of the depth estimator. Furthermore, -46.0% and -37.9% average synthesis BD-Rate gains are achieved compared to the classical approach with depth maps estimated at the encoder.

Many computer vision applications heavily rely on feature detection, description, and matching. Feature detectors are mainly based on specific image gradient distributions, which have local or global invariance to possible image translation, rotation, or to scale or affine transformation. The identifiability and invariance of features description are critical in feature matching.

In collaboration with Xi'an University (Prof. Zhaolin Xiao), we have proposed novel feature descriptors for light fields computed on the Fourier disparity layer representations of the light field (see Fig.(4)). A first feature extraction taking advantage of both the Harris feature detector and the SIFT descriptor has been proposed in 22. We have then developed a second feature descriptor, called FDL-HCGH feature, which is based on the Harris detection in a scale-disparity space, and a circular gradient histogram descriptor. It is shown to yield more accurate feature matching, compared with the reference LIght Field Feature (LiFF) descriptor, with a lower computational complexity. In order to evaluate the feature matching performance with the proposed descriptor, we have generated a synthetic stereo LF dataset with ground truth matching points.

Experimental results with synthetic and real-world datasets show that our solution outperforms existing methods in terms of both feature detection robustness and feature matching accuracy.

In collaboration with Google (Phil Chou) and the Univ. of Linkoping (Ehsan Miandji), we have introduced a novel 6-D representation of plenoptic point clouds, enabling joint, non-separable transform coding of plenoptic signals defined along both spatial and angular (viewpoint) dimensions 9. This 6-D representation, which is built in a global coordinate system, can be used in both multi-camera studio capture and video fly-by capture scenarios, with various viewpoint (camera) arrangements and densities. We show that both the Region-Adaptive Hierarchical Transform (RAHT) and the Graph Fourier Transform (GFT) can be extended to the proposed 6-D representation to enable the non-separable transform coding. Our method is applicable to plenoptic data with either dense or sparse sets of viewpoints, and to complete or incomplete plenoptic data, while the state-of-the-art RAHT-KLT method, which is separable in spatial and angular dimensions, is applicable only to complete plenoptic data. The “complete” plenoptic data refers to data that has, for each spatial point, one color for every viewpoint (ignoring any occlusions), while “incomplete” data has colors only for the visible surface points at each viewpoint. We have demonstrated that the proposed 6-D RAHT and 6-D GFT compression methods are able to outperform the state-of-the-art RAHT-KLT method on 3-D objects with various levels of surface specularity, and captured with different camera arrangements and different degrees of viewpoint sparsity.

Graph-based transforms are powerful tools for signal representation and energy compaction. However, their use for high dimensional signals such as light fields poses obvious problems of complexity. To overcome this difficulty, one can consider local graph transforms defined on supports of limited dimension, which may however not allow us to fully exploit long-term signal correlation. We have developed methods to optimize local graph supports in a rate distortion sense for efficient light field compression 13. A large graph support can be well adapted for compression efficiency, however at the expense of high complexity. In this case, we use graph reduction techniques to make the graph transform feasible. We also considered spectral clustering to reduce the dimension of the graph supports while controlling both rate and complexity (see Fig.(5)) for an example of segmentation resulting from spectral clustering). We derived the distortion and rate models which are then used to guide the graph optimization. We developed a complete light field coding scheme based on the proposed graph optimization tools. Experimental results show rate-distortion performance gains compared to the use of fixed graph support. The method also provides competitive results when compared against HEVC-based and the JPEG Pleno light field coding schemes. WE also assess the method against a homography-based low rank approximation and a Fourier disparity layer based coding method.

Deep generative models have proven to be effective priors for solving a variety of image processing problems. However, the learning of realistic image priors, based on a large number of parameters, requires a large amount of training data. It has been shown recently, with the so-called deep image prior (DIP), that randomly initialized neural networks can act as good image priors without learning.

We have proposed a deep generative model for light fields, which is compact and which does not require any training data other than the light field itself. The proposed network is based on both a generative model that aims at modeling the spatial information that is static, i.e., found in all light field views, and on a convolutional Gated Recurrent Unit (ConvGRU) that is used to model variations between angular views. The spatial view generative model is inspired from the deep decoder, itself built upon the deep image prior, but that we enhance with spatial and channel attention modules, and with quantization-aware learning. The attention modules modulate the feature maps at the output of the different layers of the generator. In addition, we offer an option which expressively encodes the upscaling operations in learned weights in order to better fit the light field to process. The deep decoder is also adapted in order to model several light field views, with layers (i.e. features) that are common to all views and others that are specific to each view. The weights of both the convGRU and the deep decoder are learned end-to-end in order to minimize the reconstruction error of the target light field.

To show the potential of the proposed generative model, we have developed a complete light field compression scheme with quantization-aware learning and entropy coding of the quantized weights. Experimental results show that the proposed method outperforms state-of-the-art light field compression methods as well as recent deep video compression methods in terms of both PSNR and MS-SSIM metrics.

State-of-the-art 2D image compression schemes rely on the power of convolutional neural networks (CNNs). Although CNNs offer promising perspectives for 2D image compression, extending such models to omnidirectional images is not straightforward. First, omnidirectional images, when compressed on 2D maps, have specific spatial and statistical properties that can not be fully captured by current CNN models. Second, basic mathematical operations composing a CNN architecture, e.g., translation and sampling, are not well-defined when the omnidirectional image is defined directly on the sphere. We have studied the learning of representation models for on-the-sphere omnidirectional images and we have proposed to use the properties of HEALPix uniform sampling of the sphere to redefine the mathematical tools used in deep learning models for omnidirectional images. In particular, i) we have proposed the definition of a new convolution operation on the sphere that keeps the high expressiveness and the low complexity of a classical 2D convolution; ii) we have adapted standard CNN techniques such as stride, iterative aggregation, and pixel shuffling to the spherical domain; and then iii) we have applies our new framework to the task of omnidirectional image compression. Our experiments shown that our proposed on-the-sphere solution leads to a better compression gain that can save 13.7% of the bit rate compared to similar learned models applied to equirectangular images. Also, compared to learning models based on graph convolutional networks, our solution supports more expressive filters that can preserve high frequencies and provide a better perceptual quality of the compressed images. Such results demonstrate the efficiency of the proposed framework, which opens new research venues for other omnidirectional vision tasks to be effectively implemented on the sphere manifold.

On-the-sphere compression of omnidirectional videos is a very promising approach. First, it saves computational complexity as it avoids to project the sphere onto a 2D map, as classically done. Second, and more importantly, it allows to achieve a better rate-distortion tradeoff, since neither the visual data nor its domain of definition are distorted. In 20, the on-the-sphere compression for omnidirectional still images, previously developed, is extended to videos. We have first proposed a complete review of existing spherical motion models. Then we have proposed a new one called tangent-linear+t. We have finally proposed a rate-distortion optimized algorithm to locally choose the best motion model for efficient motion estimation/compensation. For that purpose, we have additionally proposed a finer search pattern, called spherical-uniform, for the motion parameters, which leads to a more accurate block prediction. The novel algorithm leads to rate-distortion gains compared to methods based on a unique motion model.

In the context of the Lichie project, in collaboration with Airbus, we address two problems for satellite imaging: quasi-lossless compression and restoration, using deep learning methods.

More precisely, we developed an end-to-end trainable neural network for satellite image compression. The proposed approach builds upon an image compression scheme based on variable autoencoders with a learned hyper-prior that captures dependencies in the latent space for entropy coding. We explore this architecture in light of specificities of satellite imaging: processing constraints on board the satellite (complexity and memory constraints) and quality needed in terms of reconstruction for the processing task on the ground. We explored data augmentation to improve the reconstruction of challenging image patterns. The proposed model outperforms the current standard of lossy image compression onboard satellite-based on JPEG 2000, as well as the initial hyperprior architecture designed for natural images.

In parallel, we have developed a method to estimate the components of the Retonex model using untrained deep generative networks to restore low light satellite images. The Retinex model has indeed been shown to be an effective tool for low-light image restoration. This model assumes that an image can be decomposed into a product of two components, the illumination and the reflectance. Efficient methods have been proposed to estimate these components based on deep neural networks trained in a supervised manner with a dataset of paired low/normal-light images. However, collecting these samples is extremely challenging in practice. The proposed approach does not require any training data other than the input low-light image. To demonstrate the efficiency of the proposed estimation method, we perform simple gamma corrections on the illumination and reflectance components. We show that our approach leads to better restoration results than existing unsupervised methods and on par with fully supervised solutions thanks to the decomposition process.

In the context of the Cifre contract with Ateme, we investigate deep learning architectures for the inference of coding modes in video compression algorithms with the ultimate goal of reducing the encoder complexity. In particular, we studied the recently finalized video compression standard VVC. Compared to its predecessor standard HEVC, VVC offers about 50% compression efficiency gain, in terms of rate, at the cost of about 10x more encoder complexity. We therefore constructed a CNN-based method to speed up the partitioning of an image into blocks. More precisely, an image is first split into fixed-size so called coding tree unit. Then, each CTU is partitioned into blocks called CU which are adapted to the content. This operation, being adapted to the content, is of a extreme computational complexity, as it requires to perform for each possible partition, the whole encoding and its Rate-distortion optimization. The proposed CNN allows to avoid to test partitions that are unlikely to be selected. Thanks to a light-weight CNN, experiments show that the proposed method can achieve acceleration ranging from 17% to 35% with a reasonable efficiency drop ranging from 0.32% to 1.21% in terms of rate.

In the context of the Cifre contract with MediaKind, we develop coding tools in order to compress and deliver video, while adapting the quality to the available bandwidth and/or the user screen resolution. As a first step towards this goal, we studied in 19 bitrate ladders for the last standardized video coder, named Versatile Video Coder (VVC). Indeed, many video service providers take advantage of bitrate ladders in adaptive HTTP video streaming to account for different network states and user display specifications by providing bitrate/resolution pairs that best fit client's network conditions and display capabilities. These bitrate ladders, however, differ when using different codecs and thus the couples bitrate/resolution differ as well. In addition, bitrate ladders are based on previously available codecs (H.264/MPEG4-AVC, HEVC, etc.), i.e. codecs that are already in service, hence the introduction of new codecs e.g. VVC requires re-analyzing these ladders. For that matter, we analyzed the evolution of the bitrate ladder when using VVC. We showed how VVC impacts this ladder when compared to HEVC and H.264/AVC and in particular, that there is no need to switch to lower resolutions at the lower bitrates defined in the Call for Evidence on Transcoding for Network Distributed Video Coding (CfE).

We have pursued our development of a learning-based framework for light field view synthesis from a subset of input views, for which we published preliminary results at CVPR 2020. We have in particular proposed a deep residual architecture that can be used both for synthesizing high quality angular views in light fields and temporal frames in classical videos. The proposed framework consists of an optical flow estimator optimized for view synthesis, a trainable feature extractor and a residual convolutional network for pixel and feature-based view reconstruction. Among these modules, the fine-tuning of the optical flow estimator specifically for the view synthesis task yields scene depth or motion information that is well optimized for the targeted problem. In cooperation with the end-to-end trainable encoder, the synthesis block employs both pixel-based and feature-based synthesis with residual connection blocks, and the two synthesized views are fused with the help of a learned soft mask to obtain the final reconstructed view. Experimental results with various datasets show that our method performs favorably against other state-of-the-art methods with a large gain for light field view synthesis. Furthermore, with a little modification, our method can also be used for video frame interpolation, generating high quality frames compared with existing interpolation methods. We have also proposed a specific deep learning-based network for video frame rate up-conversion (or video frame interpolation)in 21. The proposed optical flow-based pipeline employs deep features extracted to learn residue maps for progressively refining the synthesized intermediate frames 24.

We have also proposed a learning-based method to extrapolate novel views from axial volumes of sheared epipolar plane images (EPIs), which allows us to increase the axial light field resolution 14. Axial light field resolution refers to the ability to distinguish features at different depths by refocusing. With the proposed method, the extrapolated light field gives re-focused images with a shallower depth of field (DOF), leading to more accurate refocusing results. The refocusing precision can be essential for some light field applications like microscopy. The proposed approach does not need accurate depth estimation. Experimental results with both synthetic and real light fields show that the method not only works well for light fields with small baselines as those captured by plenoptic cameras, but also applies to light fields with larger baselines.

Finally, we have designed s lightweight neural network architecture with an adversarial loss for generating a full light field from one single image 17. The method is able to estimate disparity maps and automatically identify occluded regions from one single image thanks to a disparity confidence map based on forward-backward consistency checks. The disparity confidence map also controls the use of an adversarial loss for occlusion handling. The approach outperforms reference methods when trained and tested on light field data. Besides,we also designed the method so that it can efficiently generate a full light field from one single image, even when trained only on stereo data. This allows us to generalize our approach for view synthesis to more diverse data and semantics 23.

Recent methods have been introduced with the goal of combining the advantages of well understood iterative optimization techniques with those of learnable complex image priors. A first category of methods, referred to as ”Plug-and-play” methods, has been introduced where a learned network-based prior is plugged in an iterative optimization algorithm. These learnable priors can take several forms, the most common ones being: a projection operator on a learned image subspace, a proximal operator of a regularizer or a denoiser.

In the context of the AI chair DeepCIM, we have first studied Plug-and-Play optimization for solving inverse problems by plugging a denoiser into a classical optimization algorithm. The denoiser accounts for the regularization and therefore implicitly determines the prior knowledge on the data, hence replacing typical handcrafted priors. We have extended the concept of plug-and-play optimization to use denoisers that can be parameterized for non-constant noise variance. In that aim, we have introduced a preconditioning of the ADMM algorithm, which mathematically justifies the use of such an adjustable denoiser. We additionally proposed a procedure for training a convolutional neural network for high quality non-blind image denoising that also allows for pixel-wise control of the noise standard deviation. We have shown that our pixel-wise adjustable denoiser, along with a suitable preconditioning strategy, can further improve the plug-and-play ADMM approach for several applications, including image completion, interpolation, demosaicing and Poisson denoising. An illustration of Poisson denoising results is given in Fig.(7).

One advantage of such learned priors is their genericity in the sense that they can be used for any inverse problem, and do not need to be re-trained for each new problem, in contrast with deep models learned as a regression function for a specific task. However, priors learned independently of the targeted problem may not yield the best solution. Unrolling a fixed number of iterations of optimization algorithms is another way of coupling optimization and deep learning techniques. The learnable network is trained end to- end within the iterative algorithm so that performing a fixed number of iterations yields optimized results for a given inverse problem. Several optimization algorithms (Iterative Shrinkage Thresholding Algorithm (ISTA), Half S Quadratic Splitting (HQS), and Alternating Direction Method of Multipliers (ADMM)) have been unrolled in the literature, where a learned regularization network is used at each iteration of the optimization algorithm.

While usual iterative methods iterate until idempotence, i.e. until the difference between the input and the output is sufficiently small, the number of iterations in unrolled optimization methods is set to a small value. This makes it possible to learn a component end-to-end within the optimization algorithm, hence in a way which takes into account the data term, i.e., the degradation operator. But learning networks end-to-end within an unrolled optimization scheme requires high GPU memory usage since the memory used for the backpropagation scales linearly with the number of iterations. This explains why the number of iterations used in an unrolled optimization method is limited. To cope with these limitations, we have developed a stochastic implicit unrolled proximal point algorithm with a learned denoiser, in which sub-problems are defined per iteration. We exploit the fact that the Douglas-Rachford algorithm is an application of the proximal point algorithm to re-define the unrolled step as a proximal mapping. We focused on the unrolled ADMM, which has been demonstrated to be a special case of the Douglas- Rachford algorithm, hence of the proximal point algorithm. This allows us to introduce a novel unrolled proximal gradient method coupling an implicit model and a stochastic learning strategy. We have shown that this stochastic iteration update strategy better controls the learning at each unrolled optimization step, hence leads to a faster convergence than other implicit unrolled methods, while maintaining the advantage of a low GPU memory usage, as well as similar reconstruction quality to the best unrolled methods for all considered image inverse problems.

We have also considered untrained generative model and proposed an optimization method coupling a learned denoiser with the untrained generative model, called deep image prior (DIP) in the framework of the Alternating Direction Method of Multipliers (ADMM) method 18. We have also studied different regularizers of DIP optimization, for inverse problems in imaging, focusing in particular on denoising and super-resolution. The goal was to make the best of the untrained DIP and of a generic regularizer learned in a supervised manner from a large collection of images. When placed in the ADMM framework, the denoiser is used as a proximal operator and can be learned independently of the considered inverse problem. We show the benefits of the proposed method, in comparison with other regularized DIP methods, for two linear inverse problems, i.e., denoising and super-resolution

In the Intercom project, we have studied the impact of interactivity on the coding performance. We have, for example, tackled the following problem: is it possible to compress a 360-degree video once for all, and then partly extract and decode what is needed for a user navigation, while, keeping good compression performance? First, we derived the achievable theoretical bounds in terms of storage and transmission rates. In 15, we analyzed and improved a practical coding scheme. We considered a binarized coding scheme, which insures a low decoding complexity. First, we showed that binarization does not impact the transmission rate but only slightly the storage with respect to a symbol based approach. Second, we proposed a Q-ary symmetric model to represent the pairwise joint distribution of the sources instead of the widely used Laplacian model. Third, we introduced a novel pre-estimation strategy, which allows to infer the symbols of some bit planes without any additional data and therefore permits to reduce the storage and transmission rates. In the context of 360

Previously, we derived information theoretical bounds of the compression problem with interactivity under a vanishing error probability assumption, a classical framework in information theory. In practical systems however, the core algorithm (entropy coder) needs to achieve exactly zero-error for any blocklength. Therefore, to complete our previous work, it was of great importance to also derive the compression performance in a zero-error framework. To do so, we modeled in 16 the interactive compression problem as a source coding problem when side-information (SI) may be present. Indeed, the side information may represent an image that could have been requested previously by the user. In particular, we showed that both zero-error and vanishing error schemes achieve exactly the same asymptotic compression performance. The proof technique relied on a random coding argument, and a code construction based on coset partitioning obtained from a linear code.

In the Intercom project, after deriving the achievable compression performance, we have built a new omnidirectional video coder. This original architecture enables to reduce significantly the cost of interactivity compared to the conventional video coders. In the project ICOV, we are developing a complete and well specified coder that is aimed to be shared with the community, starting from this promising proof of concept. In the year 2021, we have first worked on the bitstream specification. As for the video standards, we have defined the structure of the binary code that is stored on the servers and transmitted to the decoder. Then, we have worked on the implementation of the hierarchical channel coder that is the new entropy coding strategy that enables flexible decoding. We have tested the performance and compared them to the theoretical Shannon entropy, demonstrating the small gap between the theoretical results and the one achieved in practice.

Compression algorithms are nowadays overwhelmed by the tsunami of visual data created everyday. Despite a growing efficiency, they are always constrained to minimize the compression error, computed in the pixel domain.

The Data Repurposing, proposed in the team, reinvents how compression is done. It consists in semantically describing the database information in a concise representation, thus leading to drastic compression ratios exactly as a music score is able to describe, for example, a concert in a compact and reusable form. This enables the compression to withdraw tremendous amount of useless, or at least not essential, information while condensing the important information into a compact recycled signal. In a nutshell, in the Data Repurposing framework, the decoded signals target subjective exhaustiveness of the information
description, rather than fidelity to the input data, as in the traditional compression algorithms. In the exploratory action DARE, we had the chance to explore two directions.

We first introduce the concept of perceived information (PI), which reflects the information perceived by a given user experiencing a data collection, and which is evaluated as the volume spanned by the sources features in a personalized latent space. We use this PI metric in order to formalize a database sampling algorithm. The goal is to take into account the user's preferences while keeping a certain level of diversity in the sampled database (Figure 8). A first version of our algorithm outperforms benchmark solutions with simulation results, showing the gain in taking into account users' preferences while also maximizing the perceived information in the feature domain. We are currently working on the extension of such algorithm for real images as inputs.

A second direction for Data Repurposing explored this year deals with the generative compression. Basically, it consists in allowing the compression algorithm to “reinvent” part of the data at the decoding phase, and thus saving a lot of bit-rate by not coding it. This work is currently at a preliminary stage. We have started our study on the possibility of shaping the latent space of the compressed description such that it includes a description of the semantic. In 2021, a Young researcher ANR project has also been accepted on this research theme. It will begin in April 2022 officially.

The goal of this Cifre contract is to develop novel compression methods for 6 DoF immersive video content. This implies investigating depth estimation and view synthesis methods that would be robust to quantization noise, for which deep learning solutions are being considered. This also implies developing the corresponding coding mode decisions based on rate-distortion criteria.

The goal of this Cifre contract is to investigate deep learning architectures for the inference of coding modes in video compression algorithms with the ultimate goal of reducing the encoder complexity. The first step addresses the problem of Intra coding modes and quad-tree partitioning inference. The next step will consider Inter coding modes taking into account motion and temporal information.

The goal of this contract is to investigate deep learning methods for low light vision with sattelite imaging. The SIROCCO team focuses on two complementary problems: compression of low light images and restoration under conditions of low illumination, and hazing. The problem of low light image enhancement implies handling various factors simultaneously including brightness, contrast, artifacts and noise. We investigate solutions coupling the retinex theory, assuming that observed images can be decomposed into reflectance and illumination, with machine learning methods. We address the compression problem taking into account the processing tasks considered on the ground such as the restoration task, leading to an end-to-end optimization approach.

The goal of this study is to analyze the video compression standard recently standardized and called Versatile Video Coding (VVC) in the context of streaming. In particular, the ultimate goal is to provide the users with the best user experience in other words the best tradeoff between rate and complexity taking into account the bandwidth limitation at the user side. This optimization is performed while adjusting both the resolution of the video and the quantization level, and the optimization result is given in terms of a curve called bitrate-ladder and provides for each user bandwidth rate, the best video encoder configuration (resolution quantization).

The goal of this Cifre contract is to optimize a streaming solution taking into the whole process, namely the encoding, the long-term and the short term storages (in particular for replay, taking into the popularity of the videos), the multiple copies of a video (to adapt to both the resolution and the bandwidth of the user), and the transmissions (between all entities: encoder, back-end and front-end server, and the user). This optimization will be with several objectives as well. In particular, the goals will be to maximize the user experience but also to save energy and/or the deployment cost of a streaming solution.

All imaging systems, when capturing a view, record different combinations of light rays emitted by the environment. In a conventional camera, each sensor element sums all the light rays emitted by one point over the lens aperture. Light field cameras instead measure the light along each ray reaching the camera sensors and not only the sum of rays striking each point in the image. In one single exposure, they capture the geometric distribution of light passing through the lens. This process can be seen as sampling the plenoptic function that describes the intensity of the light rays interacting with the scene and received by an observer at every point in space, along any direction of gaze, for all times and every wavelength.

The recorded flow of rays (the light field) is in the formof high-dimensional data (4D or 5D for static and dynamic light fields). The 4D/5D light field yields a very rich description of the scene enabling advanced creation of novel images from a single capture, e.g. for computational photography by simulating a capture with a different focus and a different depth of field, by simulating lenses with different apertures, by creating images with different artistic intents. It also enables advanced scene analysis with depth and scene flow estimation and 3D modeling. The goal of the ERC-CLIM project is to develop algorithms for the entire static and video light fields processing chain. The planned research includes the development of:

Plenoptic Imaging (PLENOPTIMA) is a four-year (2021–2024) H2020 Marie Sklodowska-Curie Innovative Training Network that develops a cross-disciplinary approach to plenoptic imaging, which includes new optical materials and sensing principles, signal processing methods, new computing architectures, and vision science modelling. The ultimate goal of PLENOPTIMA is to establish new cross-sectorial, international, multi-university sustainable doctoral degree programmes in the area of plenoptic imaging and to train fifteen next generation researchers and creative professionals within these programmes for the benefit of a variety of application sectors.

Lossy compression algorithms trade bits for quality, aiming at reducing as much as possible the bitrate needed to represent the original source (or set of sources), while preserving the source quality. In the exploratory action "DARE", we propose a novel paradigm of compression algorithms, aimed at minimizing the information loss perceived by the final user instead of the actual source quality loss, under compression rate constraints.
In particular, we plan to measure the amount of information spanned by a data collection in the semantic domain. First, it enables to identify the high-level information contained in each of the image/video of a data collection. Second, it permits to take into account the redundancies and dissmilarities in the calculation of the global volume of information that is contained in a data collection. Finally, we propose to take into account the user's preferences in this calculation, since two users may have different tastes and priorities.
Once the measure of information is set, we plan to build efficient sampling algorithms to reduce the data collection's size.
This project also enables to explore new ideas for image generative compression, when part of the content can be “invented” at the decoder side.

The project aims at leveraging recent advances in three fields: image processing, computer vision and machine (deep) learning. It will focus on the design of models and algorithms for data dimensionality reduction and inverse problems with emerging image modalities. The first research challenge will concern the design of learning methods for data representation and dimensionality reduction. These methods encompass the learning of sparse and low rank models, of signal priors or representations in latent spaces of reduced dimensions. This also includes the learning of efficient and, if possible, lightweight architectures for data recovery from the representations of reduced dimension. Modeling joint distributions of pixels constituting a natural image is also a fundamental requirement for a variety of processing tasks. This is one of the major challenges in generative image modeling, field conquered in recent years by deep learning. Based on the above models, our goal is also to develop algorithms for solving a number of inverse problems with novel imaging modalities. Solving inverse problems to retrieve a good representation of the scene from the captured data requires prior knowledge on the structure of the image space. Deep learning based techniques designed to learn signal priors, tcan be used as regularization models.

This project aims to secure the industrial impact of the InterCom project, ended in Dec. 2020, and funded by the Labex CominLabs. Indeed, the goal of the Cominlabs InterCom project was to design novel compression algorithms to allow interactive communication between users and a server. One instance of this interactive scenario is a visual immersive experience, where a user can navigate freely in a 3D scene. This leads to tremendous amount of visual data, such that the whole scene cannot be sent to the user. Fortunately, the user watches only a part of the scene, such that only the requested part needs to be sent to the user. This however presents a great challenge from the point of view of the compression. Indeed, the data needs to be compressed once online, but decompressed in many manners, one manner being one requested point of view. In the case of a static 360 camera, the navigation has 3 Degrees of Freedom such that there is one decompression per value of a 3D-vector. One of the achievement of the InterCom project is a prototype for an interactive compression scheme for 360-degree images. The MOVE project helps the ADT ICOV project towards the construction of a full demonstrator. It also helps the maturation of the startup project lead by Navid Mahmoudian Bidgoli (ANAX).

The amount of data available online is growing so fast that it is essential to rely on advanced Machine Learning techniques so as to automatically analyze, sort, and organize the content uploaded by e.g. sensors or users. The conventional data transmission framework assumes that the data should be completely reconstructed, even with some distortions, by the server. Instead, this project aims to develop a novel communication framework in which the server may also apply a learning task over the coded data. The project will therefore develop an Information Theoretic analysis so as to understand the fundamental limits of such systems, and develop novel coding techniques allowing for both learning and data reconstruction from the coded data.

Two former PhD students of the team (Navid Mahmoudian Bidgoli and Simon Evain) plan to launch a startup, named Anax, on the theme of omnidirectional/image processing. They have obtained a one year grant (Sept 2021 - Aug 2022) for both of them funded by the Inria Startup Studio. Here is the description of the Anax project.

Anax aims to provide a deep tech software solution for processing 360-degree visual content with artificial intelligence (AI) specially designed for the preparation of virtual tours. Anax is aimed at various actors who wish to offer an immersive visit experience to improve their visibility, such as real estate agencies, cultural institutions, and interior designers. Anax is developing a technology that allows retrieving a faithful 3D reconstruction of a building from 360-degree images, opening up a wide range of applications based on AI image processing such as automatic recommendation of similar apartments, augmented reality, automatic inventory, etc. The envisaged solution, based on artificial intelligence, works even with consumer 360-degree capturing devices that are readily accessible to the general public.