2025Activity report​Project-TeamCOMPACT

Axis 1.1: Compression adapted​​​‌ to the data-type

The‌ field of visual data‌​‌ compression knows new challenges​​​‌ triggered by the emergence​ of novel modalities (light​‌ fields, aka plenoptic ,​​ 360o videos, and​​​‌ even holographic data). This​ research axis focuses on​‌ compact representation of light​​ fields. Unlike traditional cameras​​​‌ which capture simple 2D​ images, light field cameras​‌ capture very large volumes​​ of high-dimensional data containing​​​‌ information about the light​ rays as they interact​‌ with the physical objects​​ in the scene. A​​​‌ major challenge in the​ practical use of light​‌ field technology is the​​ huge amount of captured​​​‌ data, hence the need​ for efficient compression solutions.​‌ While in the past​​ decade the problem has​​​‌ been addressed using traditional​ signal processing models, e.g.​‌ sparse or low rank​​ models, these models present​​​‌ some limitations in terms​ of well capturing and​‌ representing the characteristics of​​ real data. Real data​​​‌ in general require much​ more complex models that​‌ cannot be fully expressed​​ analytically. By contrast, machine​​​‌ learning (ML) methods are​ data-driven approaches which, by​‌ learning a very large​​ number of parameters, turn​​​‌ out to be more​ powerful for encoding and​‌ expressing complex data properties.​​ This is especially important​​​‌ for plenoptic data which​ represents the complexity of​‌ the visual worlds in​​ terms of reflective, diffusive,​​​‌ semi-transparent and partially-occluded objects​ at various depths. In​‌ this context, this research​​ axis aims at dealing​​​‌ with high dimensional light​ field data, focusing on​‌ problems of dimensionality reduction​​ for compression while enabling​​​‌ rendering of high quality.​ Another problem that will​‌ be investigated corresponds to​​ the case where the​​​‌ light field or plenoptic​ data is first represented​‌ by a deep network​​ model. The problem of​​​‌ data compression then becomes​ a problem of dimensionality​‌ reduction of Deep Network​​ Models, e.g. for Mobile​​​‌ Computational Plenoptics.

Axis 1.2:​ Compression adapted to the​‌ user-type: Machine

The volumes​​ of visual data being​​​‌ generated 40 are such​ that these data will​‌ not only be viewed​​ by humans but also​​​‌ by machines. For instance,​ in autonomous vehicles, the​‌ machine is the perception​​ system that processes videos​​​‌ to detect objects such​ as pedestrians, vehicles, traffic​‌ signs, and barriers. Another​​ example is the case​​​‌ when a tremendous amount​ of visual data is​‌ uploaded (in social media​​ for instance) and analyzed​​​‌ to make recommendations to​ humans. A notable difference​‌ between compression for humans​​ and compression for machines​​​‌ is that in the​ case of machines the​‌ entirety of the image​​ is not necessary but​​​‌ only some elements are​ needed to perform the​‌ analysis. Hence there is​​ a need to develop​​​‌ specific algorithms for compression​ for machines.

Furthermore, among​‌ the use cases of​​ compression for machines, we​​​‌ can distinguish two scenarios.​ In the case of​‌ cameras embedded in autonomous​​ vehicles, it is known,​​​‌ upon acquisition, that these​ visual data will be​‌ destined for machines. However,​​ due to time and/or​​​‌ computational constraints, the analysis​ cannot be performed at​‌ the camera, and the​​ data need to be​​​‌ compressed and sent to​ a remote machine. Instead,​‌ in the second example​​ of data uploaded on​​ a social media, the​​​‌ primary destination of the‌ data was initially a‌​‌ human, but it is​​ later decided, after compression,​​​‌ that these data will‌ be analyzed by a‌​‌ machine. For these two​​ use cases, the challenges​​​‌ are different. In the‌ first case, the challenge‌​‌ is to (i) develop​​ new compression algorithms that​​​‌ take into account the‌ receiver, machine and the‌​‌ task that will be​​ performed. In the second​​​‌ case, the goal is‌ instead to (ii) develop‌​‌ algorithms that process the​​ data directly in the​​​‌ compressed domain when the‌ compression algorithm has been‌​‌ specifically designed for human​​ vision.

To develop new​​​‌ compression algorithms (i), our‌ approach is to first‌​‌ define the achievable compression​​ rates when the receiver​​​‌ is a machine that‌ is not interested in‌​‌ the entirety of the​​ data but aims to​​​‌ perform processing on it.‌ Our approach will differ‌​‌ from the work of​​ the community 41,​​​‌ 44, 45 in‌ that we incorporate a‌​‌ strict guarantee on the​​ quality of the processing​​​‌ output. The long term‌ objective is to design‌​‌ compression algorithms, where the​​ task may not be​​​‌ known in advance or‌ another task may be‌​‌ chosen (for instance, a​​ new category to be​​​‌ detected).

When the objective‌ is to build algorithms‌​‌ that allow for processing​​ compressed data with an​​​‌ existing algorithm primarily designed‌ for humans (ii), our‌​‌ approach is to avoid​​ decompressing the data. By​​​‌ avoiding data decompression, it‌ is possible to work‌​‌ with more compact representations​​ of the data. The​​​‌ community avoids this decompression‌ when compression is learned‌​‌ for a specific task​​ (i), as in 64​​​‌, 36, 37‌. Conversely, our objective‌​‌ is to construct these​​ algorithms when the compression​​​‌ is performed by an‌ existing algorithm intended for‌​‌ human viewers.

Axis 1.3:​​ Compression adapted to the​​​‌ media

Axis​​​‌ 2.1: Ultra-low bitrate visual‌ data compression

The goal‌​‌ of this axis is​​ to reduce the storage​​​‌ cost of cold data,‌ by achieving very high‌​‌ compression ratio. Recently, researchers​​ have proven the existence​​​‌ of a trade-off between‌ distortion and perception when‌​‌ compressing data at low​​ bitrate32. In​​​‌ other words, targeting low‌ bitrate inevitably leads to‌​‌ move away from the​​ traditional compression's objective, i.e.,​​​‌ keeping faithful decoded data,‌ and to target visual‌​‌ plausibility instead. Therefore, the​​ envisaged solution will semantically​​​‌ describe the visual 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 semantic description.​​ At the decoder side,​​​‌ a generative process, relying‌ for example on Diffusion‌​‌ Models 60, is​​ in charge of reconstructing​​​‌ the image or video‌ that is close semantically‌​‌ to the input. In​​ a nutshell, the decoded​​​‌ signals target subjective exhaustiveness‌ of the information description,‌​‌ rather than fidelity to​​ the input data, as​​​‌ in the traditional compression‌ algorithms. Naturally, not all‌​‌ the visual content is​​ meant to be regenerated.​​​‌ Users might be willing‌ to retrieve faithfully the‌​‌ content after decompression. Such​​ approaches will therefore be​​​‌ designed according to user’s‌ profile taking into account‌​‌ their choice and interaction.​​ This is a complete​​​‌ change of paradigm, which‌ must enable gigantic compression‌​‌ gains. Considering this approach​​ would use heavy deep​​​‌ learning algorithms and may‌ not tackle data that‌​‌ are often decoded, otherwise​​ the energy due to​​​‌ storage cost reduction would‌ be totally negligible when‌​‌ compared with the huge​​ decoding complexity. On the​​​‌ contrary, this would perfectly‌ fit with cold data.‌​‌ Finally, in order to​​ be coherent with the​​​‌ purpose of sobriety, we‌ will look for solutions‌​‌ that do not require​​ retraining or even fine-tuning​​​‌ of the heavy Diffusion‌ Models.

Axis 2.2: Data‌​‌ collection sampling

As previously​​ stated, the amount of​​​‌ data created every day‌ is huge and exploding.‌​‌ This is certainly accelerated​​ by the fact that​​​‌ most of the social‌ network, video platforms or‌​‌ mobile companies offer the​​ possibility to create, stream​​​‌ and store unlimited data‌ size (or with unreachable‌​‌ bounds), leaving the impression​​​‌ that the storage of​ data is intangible and​‌ cost-less in terms of​​ energy consumption. Increasing the​​​‌ awareness of users or​ companies requires an efficient​‌ way to automatically decide​​ what data deserves to​​​‌ be kept or deleted.​

In this axis, we​‌ will explore data collection​​ sampling, which consists in​​​‌ selecting the images and​ videos a user would​‌ like to keep among​​ a massive data collection,​​​‌ enabling significant data size​ savings. This requires first​‌ modeling the information perceived​​ by a given user​​​‌ when experiencing a data​ collection (the initial or​‌ the sampled one). This​​ model relies on the​​​‌ volume spanned by the​ sources features in a​‌ personalized latent space. In​​ parallel, we will develop​​​‌ methods to learn the​ structure and statistics that​‌ rule a given data​​ collection. Concretely, among all​​​‌ the pictures of an​ image collection, some coherent​‌ patterns (e.g.,​​ landscape, portrait), resemblance between​​​‌ images, chronological landscape evolution​ or any salient content​‌ can be learned and​​ described by mathematical tools,​​​‌ for example with graphs​ or manifolds. Thereafter these​‌ structures will be the​​ support of sampling algorithms​​​‌ aiming at the subjective​ exhaustiveness of the description,​‌ i.e., covering the​​ maximum volume of the​​​‌ learned structure. We will​ thus pose the trade-off​‌ between the rate of​​ the samples (not necessarily​​​‌ taken from the input​ data, but could be​‌ a combination of them)​​ and the quality of​​​‌ the obtained description, driven​ by the user’s preferences.​‌

Axis 2.3: Low-tech video​​ coders

All the recent​​​‌ advances in video compression​ are due to an​‌ increase of the complexity:​​ e.g., more tools and​​​‌ more freedom in the​ choice of parameters 34​‌ or fully deep learning-based​​ algorithms 55. In​​​‌ such a context, the​ global energy cost due​‌ to video consumption can​​ only explode, which is​​​‌ not compatible with the​ urgent need of energetic​‌ sobriety. Developing low-energetic video​​ compression/decompression algorithms has been​​​‌ explored for a long​ time 51, 29​‌, 59. However,​​ most of the time,​​​‌ the achieved low complexity​ of the compression algorithms​‌ comes from the reduction​​ of the capability of​​​‌ the video coder (​e.g., less parameters to​‌ estimate, removing of some​​ complex functionalities). Such approaches​​​‌ do not put in​ question the trade-off between​‌ complexity and video coding​​ performance, and thus remain​​​‌ limited.

In this axis,​ we plan to investigate​‌ low complexity algorithms that​​ are not low-cost versions​​​‌ of a complex algorithm.​ The proposed methodology is​‌ the following. We start​​ from a complex learning-based​​​‌ coder as for example​ the auto-encoder-like architecture proposed​‌ in 54. Such​​ architectures are able to​​​‌ achieve outstanding performance, with,​ however a gigantic encoding​‌ and decoding complexity. Our​​ goal is to investigate​​​‌ how to deduce from​ this trained network and​‌ its millions of parameters,​​ some efficient features for​​​‌ low complexity compression. As​ an example, we can​‌ show that the set​​ of non-linear operations involved​​​‌ in a deep convolutional​ neural architecture can be​‌ modeled as a linear​​ operation once the input​​ is fixed, like​​​‌ it is studied in‌ 57, 58.‌​‌ The strength of the​​ deep architecture resides in​​​‌ its ability to adjust‌ this linear filter to‌​‌ the input. For our​​ purpose, we will, on​​​‌ the contrary, investigate if‌ some common features reside‌​‌ in these linear filters​​ when the input is​​​‌ changed. These common features‌ may constitute, for example,‌​‌ an efficient transform or​​ partitioning operation that does​​​‌ not require anymore millions‌ of parameters. In a‌​‌ nutshell, the intuition will​​ be to take benefit​​​‌ of algorithms trained on‌ a large set of‌​‌ images and to extract​​ from them some common​​​‌ analysis tools.

Axis 2.4:‌ Sobriety in video usage‌​‌

Rebound effect or the​​ Jevons’s paradox 67 refers​​​‌ to the fact that‌ reducing the cost (in‌​‌ terms of energy or​​ resource consumption) of a​​​‌ technology often leads to‌ an increase of the‌​‌ technology usage and thus​​ to a global increase​​​‌ of the cost, in‌ opposition with the initial‌​‌ goal. Video compression is​​ clearly a good example​​​‌ of this rebound effect.‌ Smaller video sizes (and‌​‌ other technology advances) have​​ led to a global​​​‌ increase of the video‌ usage in today’s society.‌​‌ As the ultimate goal,​​ for achieving IPCC objectives,​​​‌ is to reduce the‌ global carbon footprint of‌​‌ video usage, compression nowadays​​ should not only focus​​​‌ on the reduction of‌ each video file individually.‌​‌ The compression problem should​​ be formulated globally. This​​​‌ inevitably raises the following‌ research question: what is‌​‌ the best (most efficient​​ and acceptable) solution for​​​‌ reducing the amount of‌ videos created/stored/consumed? This question‌​‌ naturally includes the study​​ of user's behavior, and​​​‌ thus deals with other‌ research fields in human‌​‌ and social sciences. The​​ goal of the team​​​‌ COMPACT is twofolds: i)‌ to raise a multidisciplinary‌​‌ research effort on that​​ question by connecting different​​​‌ laboratories and ii) to‌ put its expertise in‌​‌ video compression to the​​ service of this crucial​​​‌ question.

Axis 3.1:‌ Joint optics/processing

Our goal‌​‌ is the design of​​ an end-to-end optimization framework​​​‌ designed for acquiring high-resolution‌ images across an extensive‌​‌ Depth of Field (DOF)​​ range within a microscopy​​​‌ system. Microscopy is indeed‌ one key potential application‌​‌ of light field imaging.​​ The optics and post-processing​​​‌ algorithm will be modeled‌ as parts of the‌​‌ end-to-end differentiable computational image​​ acquisition system, allowing for​​​‌ simultaneously optimizing both components.‌ Our computational Extended DOF‌​‌ microscopy imaging system will​​ employ a hybrid approach​​​‌ combining an optical setup‌ with a learned wavefront‌​‌ modulating optical element at​​ the Fourier plane based​​​‌ on metasurfaces. The extended‌ depth of field leads‌​‌ to an increased axial​​ resolution which refers to​​​‌ the ability to distinguish‌ features at different depths‌​‌ by refocusing. While we​​ have obtained initial results​​​‌ for 2D microscopy 30‌, our goal here‌​‌ will be to extend​​​‌ these results to light​ field microscopy, which has​‌ recently retained the attention​​ of the research community​​​‌ 63, 65,​ 52.

Axis 3.2:​‌ Joint representation/processing: Neural Scene​​ Representation

The task of​​​‌ generating high-quality immersive content​ with a sufficiently high​‌ angular and spatial resolution​​ is technologically challenging, due​​​‌ to the complexity of​ the constrained capture setup​‌ and the bottleneck of​​ data storage and of​​​‌ computational cost. Reconstructing the​ imaged scene (from a​‌ few viewpoints), with a​​ sufficient resolution and quality,​​​‌ and in a way​ that we can observe​‌ it from almost continuously​​ varying positions or angles​​​‌ in space is also​ an important challenge for​‌ a wide adoption in​​ consumer applications. To address​​​‌ the two above problems,​ the concept of NeRF​‌ has been introduced as​​ an implicit model that​​​‌ maps 5D vectors (3D​ coordinates plus 2D viewing​‌ directions) to opacity and​​ color values. The model​​​‌ is based on multi-layer​ perceptrons (MLP) trained by​‌ fitting the model to​​ a set of input​​​‌ views. The learned model​ is an implicit scene​‌ representation that can be​​ used to generate any​​​‌ view of the light​ field using volume rendering​‌ techniques. A variety of​​ works have attempted to​​​‌ handle dynamic scenes in​ radiance field reconstructions but​‌ they either constrain the​​ capture process with multi-view​​​‌ or suffer from quality​ loss when compared to​‌ static scene representations. The​​ proposed research, jointly addressing​​​‌ acquisition, representation and scene​ reconstruction problems 47,​‌ 56, 43 will​​ focus on the reconstruction​​​‌ of neural radiance fields​ from a limited set​‌ of input images, especially​​ in the context of​​​‌ unconstrained, monocular captures, on​ the completion of the​‌ NeRF representation when the​​ capture is incomplete due​​​‌ to a limited set​ of input images or​‌ due to motion in​​ the scene, on the​​​‌ representations of dynamic scenes​ that are both compact​‌ (low memory) and limited​​ in computational complexity. The​​​‌ compactness of scenes will​ be explored considering joint​‌ implicit representations for a​​ collection of data points​​​‌ (2D Images or light​ fields). The implicit representations​‌ inspired from the NeRF​​ concept can be seen​​​‌ as neural network based​ data representations. The generalization​‌ of joint implicit representations​​ to unseen data points​​​‌ assumed to reside in​ the same subspace as​‌ the training data points​​ will also be investigated.​​​‌

Axis 4.1: Optimization methods‌​‌ with learned priors

Building​​ upon our past work​​​‌ aiming at taking advantage‌ of learned priors in‌​‌ optimization algorithms, i.e. via​​ plug-and-play and unrolled optimization​​​‌ methods, we will further‌ investigate Deep equilibrium (DEQ)‌​‌ 31 models. Unrolled optimization​​ methods, by coupling optimization​​​‌ algorithms with end-to-end trained‌ regularization, recently emerged as‌​‌ powerful solutions to inverse​​ problems. However, training such​​​‌ unrolled neural networks end-to-end‌ can come with a‌​‌ large memory footprint 46​​, hence their numbers​​​‌ of iterations are in‌ general limited and they‌​‌ do not generally converge.​​ DEQ models can be​​​‌ seen as an extension‌ of unrolled methods with‌​‌ a theoretically infinite amount​​ of iterations. DEQ models​​​‌ leverage fixed-point properties, allowing‌ for simpler back-propagation. We‌​‌ will further study these​​ models to learn image​​​‌ priors and apply them‌ to inverse problems in‌​‌ classical 2D and new​​ imaging modalities (light fields,​​​‌ omni-directional images).

Axis 4.2:‌ Learning on graphs

In‌​‌ the last decades, there​​ has been a multiplication​​​‌ of data that cannot‌ be properly represented by‌​‌ conventional means, but rather​​ by relationships between objects,​​​‌ of various natures and‌ with various properties. Such‌​‌ structures are usually represented​​ as graphs. This​​​‌ is for instance the‌ case of collections of‌​‌ (visual) data under the​​ form of relational databases​​​‌ (Axis 3), formed‌ by drawing “meaningful” relations‌​‌ between individual data points​​ according to some notion​​​‌ of proximity (semantic, geographical,‌ etc.). Moreover, graphs are‌​‌ increasingly used to represent​​ the structure of (potentially​​​‌ pre-trained) neural networks. Processing‌ this structure using graph‌​‌ machine learning and graph​​ signal processing tools gives​​​‌ rise to the recent‌ topic of (graph) meta-networks‌​‌49, which draws​​ connections with all other​​​‌ axes, particularly the definition‌ of low-tech encoders (Axis‌​‌ 3). Finally, graphs​​ are also a popular​​​‌ representation for geometric data‌ exhibiting invariance to certain‌​‌ transforms 33 such as​​ 2D or 3D isometries,​​​‌ often encountered in non-conventional‌ visual data (Axis 3‌​‌).

(Un)structured data such​​ as graphs posit many​​​‌ challenges. Processing and storing‌ them can be computationally‌​‌ burdensome if done naively.​​ The main challenge resides​​​‌ in the fact that‌ the regularity of other‌​‌ types of data (fixed-size​​ vectors, regular grids, well-defined​​​‌ boundaries, etc.), at the‌ basis of many methods,‌​‌ cannot be easily defined​​ here. This axis is​​​‌ thus dedicated to advancing‌ the state-of-the-art in processing‌​‌ efficiently graph data, often​​ through the lens of​​​‌ compression. ML techniques have‌ proved extremely efficient in‌​‌ designing adaptive, data-driven methods​​ for compression 66,​​​‌ including for database reduction‌ 39. Conversely, the‌​‌ extraction of information from​​ compressed databases, a fortiori​​​‌ by ML, is a‌ major requirement of any‌​‌ compression pipeline. Since graphs​​ have become the de​​​‌ facto structure to represent‌ modern relational data, graph‌​‌ ML (GML) has known​​​‌ a tremendous development in​ the last few years,​‌ with Graph Neural Networks​​ (GNN) at the forefront​​​‌ of it. Acclaimed for​ their flexibility, these deep​‌ architectures however suffer from​​ many issues, with very​​​‌ limited theoretical and empirical​ comprehension. A major goal​‌ will be to deepen​​ this understanding through the​​​‌ use of tools such​ as statistical models of​‌ large random graphs and​​ information theory. New random​​​‌ graph models adapted to​ modern real-world data will​‌ be developed, focusing on​​ databases arising from visual​​​‌ data but also generic​ databases, whose analysis will​‌ help the choice of​​ GNN architecture, and ultimately​​​‌ lead to new architecture​ improving the state-of-the-art, in​‌ terms of performance and/or​​ computational efficiency.

Axis 4.3:​​​‌ Reducing graphs

Data compression​ approaches on graphs are​‌ referred to as graph​​ reduction methods. With modern​​​‌ large graphs numbering millions​ of nodes, these methods​‌ have become a staple​​ of many pipelines, including​​​‌ ML methods mentioned above​ and database reduction (Axis​‌ 3). Graph reduction​​ can be broadly sorted​​​‌ into two related families​ of algorithms: graph sampling,​‌ and graph coarsening.

7.1.1​ color-guidance

• Keyword:
  Image compression​‌
• Scientific Description:
  This study​​ addresses the challenge of​​​‌ controlling the global color​ aspect of images generated​‌ by a diffusion model​​ without training or fine-tuning.​​​‌ We rewrite the guidance​ equations to ensure that​‌ the outputs are closer​​ to a known color​​​‌ map, without compromising the​ quality of the generation.​‌ Our method results in​​ new guidance equations. In​​​‌ the context of color​ guidance, we show that​‌ the scaling of the​​ guidance should not decrease​​​‌ but rather increase throughout​ the diffusion process. In​‌ a second contribution, our​​ guidance is applied in​​​‌ a compression framework, where​ we combine both semantic​‌ and general color information​​ of the image to​​​‌ decode at low cost.​ We show that our​‌ method is effective in​​ improving the fidelity and​​​‌ realism of compressed images​ at extremely low bit​‌ rates (0.001 bpp), performing​​ better on these criteria​​​‌ when compared to other​ classical or more semantically​‌ oriented approaches.
• Functional Description:​​
  Official implementation of the​​​‌ article: "Linearly transformed color​ guide for low-bitrate diffusion​‌ based image compression" Paper(https://arxiv.org/pdf/2404.06865)​​
• Publication:
  hal-04882103
• Contact:
  Tom​​​‌ Bordin

7.1.2 Graph coarsening​ with message-passing guarantees

• Keywords:​‌
  Graph Neural Networks, Deep​​ learning, Dimensionality reduction
• Functional​​​‌ Description:

  This repository contains​ the code for the​‌ paper “Graph coarsening with​​ message-passing guarantees”, published at​​​‌ NeurIPS 2024.

  This code​ includes Jupyter notebooks that​‌ reproduce the results (tables​​ and plots) presented in​​​‌ the paper. These experiments​ focus on using a​‌ newly proposed Propagation matrix​​ for the Graph Neural​​​‌ Network (GNN) on the​ coarsened graph.

• Publication:
  hal-04617519​‌
• Contact:
  Antonin Joly

7.1.3​​ Taxonomy of reduction matrices​​​‌ for Graph Coarsening

• Keywords:​
  Deep learning, Dimensionality reduction,​‌ Graph Neural Networks
• Functional​​ Description:

  This repository contains​​​‌ the code for the​ paper “Taxonomy of Reduction​‌ Matrices for Graph Coarsening”,​​ published at NeurIPS 2025.​​​‌

  This code includes Jupyter​ notebooks that reproduce the​‌ results (tables and plots)​​ presented in the paper.​​​‌ These experiments focus on​ optimizing reduction matrices for​‌ a fixed lifting matrix​​ in graph coarsening with​​​‌ the framework described in​ the paper.

• Publication:
  hal-05248172​‌
• Contact:
  Antonin Joly

7.1.4​​ mendevi

• Name:
  Energy measurement​​​‌ of video encoding and​ decoding
• Keywords:
  Energy, Video​‌ analysis, Video compression
• Functional​​ Description:
  1. It supports​​​‌ the libx264, libopenh264, libx265,​ libvpx-vp9, libaom-av1, libsvtav1, librav1e​‌ and vvc cpu encoders.​​ 2. It supports the​​​‌ h264_nvenc, hevc_nvenc, av1_nvenc and​ *_vaapi gpu encoders. 3.​‌ Distortions are measured using​​ the lpips, psnr, ssim,​​​‌ vif and vmaf metrics.​ 4. Complexity are measured​‌ using the rms_sobel and​​ rms_time_diff metrics. 5. Encoding​​​‌ efforts are fast, medium​ and slow. 6. It​‌ takes care about the​​ colorspaces (range, transfer and​​ primaries). 7. Iterate over​​​‌ different effort, encoder, mode,‌ quality, threads, fps, resolution‌​‌ and pix_fmt. 8. Energy​​ measurements are catched with​​​‌ RAPL and an external‌ wattmeter on grid'5000. 9.‌​‌ Get the cpu, gpu,​​ ram and temperature activity.​​​‌ 10. Get a full‌ environment context, including hardware‌​‌ and software version. 11.​​ It support the mode​​​‌ (constant bitrate) cbr and‌ (constant quality) vbr. 12.‌​‌ Ability to modify ffmpeg​​ commands on the fly​​​‌ to perform specific tests.‌ 13. It take care‌​‌ to transfer files to​​ RAM if possible to​​​‌ avoid biases related to‌ storage space access. 14.‌​‌ Provides a guide to​​ compile ffmpeg with all​​​‌ optimizations in order to‌ compare encoders/decoders at their‌​‌ limits.
• URL:
  https://­mendevi.­readthedocs.­io/­latest/
• Contact:​​
  Robin Richard

8.1.1 DUALF-D: Disentangled‌ Dual-Hyperprior Approach for Light‌​‌ Field Image Compression

Participants:​​ Soheib Takhtardeshir, Christine​​​‌ Guillemot.

Light field‌ (LF) imaging captures spatial‌​‌ and angular information, offering​​ a 4D scene representation​​​‌ enabling enhanced visual un-‌ derstanding. However, high dimensionality‌​‌ and redundancy across spatial​​ and angular domains present​​​‌ major challenges for com-‌ pression, particularly where storage,‌​‌ transmission bandwidth, or processing​​ latency are constrained. We​​​‌ have developed a novel‌ Variational Autoencoder (VAE)-based framework‌​‌ that explicitly disentangles spatial​​ and angular features using​​​‌ two parallel latent branches‌ 17, 9.‌​‌ Each branch is coupled​​ with an independent hyperprior​​​‌ model, allowing more precise‌ distribution estimation for entropy‌​‌ coding and finer rate-distortion​​ control. This dual-hyperprior structure​​​‌ enables the network to‌ adaptively compress spatial and‌​‌ angular infor- mation based​​ on their unique statistical​​​‌ characteristics, improving coding efficiency.‌ To further enhance latent‌​‌ feature specialization and promote​​ disentanglement, we introduced a​​​‌ mutual information-based regularization term‌ that minimizes redundancy between‌​‌ the two branches while​​ preserving feature diversity. Unlike​​​‌ prior methods relying on‌ covariance-based penalties prone to‌​‌ collapse, our information-theoretic regularizer​​ provides more stable and​​​‌ interpretable latent separation 8‌. Experimental results on‌​‌ publicly available LF datasets​​ demonstrate our method achieves​​​‌ strong compression performance, yielding‌ an average BD-PSNR gain‌​‌ of 2.91 dB over​​ HEVC and high compression​​​‌ ratios (e.g., 200:1). Additionally,‌ our design enables fast‌​‌ inference, with a total​​ end-to- end time over​​​‌ 19x faster than the‌ JPEG Pleno standard, making‌​‌ it well-suited for real-time​​ and bandwidth-sensitive applications. By​​​‌ jointly leveraging disentangled representation‌ learning, dual-hyperprior modeling, and‌​‌ information-theoretic regularization, our approach​​ offers a scalable, effective​​​‌ solution for practical light‌ field image compression.

8.1.2‌​‌ Zero-error information theory and​​ application to coding for​​​‌ Computing

Participants: Aline Roumy‌.

Zero-error coding encompasses‌​‌ a variety of source​​ and channel coding problems​​​‌ in which the probability‌ of error must be‌​‌ exactly zero. This requirement​​ is stricter than that​​​‌ of the classical vanishing-error‌ regime, where the error‌​‌ probability tends to zero​​ as the code blocklength​​​‌ goes to infinity. An‌ example of a zero-error‌​‌ problem is coding for​​ computing, where the goal​​​‌ is to compress data‌ not merely for visualization,‌​‌ but also to enable​​​‌ reliable inference tasks.

In​ general, zero-error coding leads​‌ to challenging open combinatorial​​ problems. In 5,​​​‌ we investigated two unsolved​ zero-error settings: the source​‌ coding problem with side​​ information and the channel​​​‌ coding problem. We focused​ on families of independent​‌ problems for which the​​ underlying probability distribution decomposes​​​‌ as a product of​ marginal distributions. A crucial​‌ step in our analysis​​ was establishing the additivity​​​‌ of the optimal rate.​ Unlike in the vanishing-error​‌ regime, this property does​​ not always hold in​​​‌ the zero-error setting. When​ additivity does hold, concatenation​‌ of optimal codes remains​​ optimal.

As a consequence,​​​‌ we derived new single-letter​ characterizations of the optimal​‌ information-theoretic rates for previously​​ unsolved graph families. In​​​‌ particular, we obtained results​ for graphs formed as​‌ products of perfect graphs​​ (which are not perfect​​​‌ in general) as well​ as for graphs obtained​‌ as the product of​​ a perfect graph and​​​‌ the pentagon graph.

8.1.3​ Coding for Machine: learning​‌ in the compressed domain​​

Participants: Rémi Piau,​​​‌ Thomas Maugey, Aline​ Roumy.

In most​‌ of the learning tasks,​​ it is necessary to​​​‌ scale the image size​ to the networks. This​‌ downsampling is generally done​​ in the pixel domain​​​‌ (it can be done​ before or inside the​‌ network itself) and thus​​ requires a decoding of​​​‌ the image at its​ full resolution which can​‌ be complex for the​​ most recent formats. Instead,​​​‌ we proposed to sample​ the image directly in​‌ the JPEG bitstream, to​​ partially decode some image​​​‌ MCU and to feed​ them to the learning​‌ task, which is challenging​​ due to the variable​​​‌ length coding involved in​ JPEG. After showing some​‌ interesting properties of the​​ JPEG bitstream, we proposed​​​‌ an end-to-end learning pipeline​ starting from a decoding​‌ of only a extracted​​ subset of the JPEG​​​‌ bitstream. Our results demonstrated​ the validity of our​‌ approach and that learning​​ directly in the JPEG​​​‌ bitstream is possible. 25​

8.1.4 Efficient Constraining of​‌ Transcoding in DNA-Based Image​​ Storage

Participants: Sara Al​​​‌ Sayyed, Aline Roumy​, Thomas Maugey.​‌

DNA has emerged as​​ a promising alternative for​​​‌ long-term data storage due​ to its high capacity,​‌ durability, and low-energy potential.​​ However, storing data in​​​‌ DNA presents several challenges.​ First, it requires complex​‌ and costly biochemical processes,​​ making efficient compression crucial​​​‌ to reducing DNA synthesis​ time and cost. Second,​‌ these processes are prone​​ to errors that must​​​‌ be avoided and/or corrected.​ In particular, homopolymers (repetitions​‌ of the same nucleotide)​​ are a well-known source​​​‌ of errors during the​ sequencing step. Avoiding such​‌ repetitions helps mitigate errors​​ but introduces a constraint​​​‌ that may increase the​ data compression rate. In​‌ this paper, we propose​​ two transcoding methods that​​​‌ address these two key​ challenges: reducing data rate​‌ and minimizing errors. The​​ first method strictly enforces​​​‌ the error-minimization constraint by​ eliminating homopolymers of a​‌ certain length, at the​​ cost of an increased​​​‌ data rate. In contrast,​ the second method accepts​‌ a slight increase in​​ homopolymers. However, we show​​ that these increases remain​​​‌ limited (2.14$\%$ increase‌ in compression rate for‌​‌ the first method and​​ 0.39$\%$ homopolymer rate​​​‌ for the second). These‌ two approaches demonstrate that‌​‌ it is possible to​​ efficiently constrain transcoding while​​​‌ balancing error minimization and‌ compression performance. This work‌​‌ was published in 10​​.

8.1.5 Compact image​​​‌ representation for content-based image‌ retrieval in DNA data‌​‌ storage

Participants: Sara Al​​ Sayyed, Aline Roumy​​​‌, Thomas Maugey.‌

In this work, we‌​‌ propose a novel image​​ compression method for content-based​​​‌ image retrieval in the‌ context of DNA data‌​‌ storage. As explained before,​​ storing data on DNA​​​‌ is an extremely promising‌ solution due to its‌​‌ compactness, long-term durability, and​​ energy efficiency. However, its​​​‌ compactness introduces two challenges:‌ the need for efficient‌​‌ data access and the​​ ability to flexibly handle​​​‌ new (and not predefined)‌ types of queries. To‌​‌ address the efficiency challenge,​​ our approach enables direct​​​‌ image retrieval within the‌ DNA domain. To ensure‌​‌ flexibility, we design a​​ compact data identifier that​​​‌ is a semantic representation‌ of the image and‌​‌ serves as a header​​ at the beginning of​​​‌ the DNA strand. Our‌ approach shows high visual‌​‌ and quantitative performance, outperforming​​ state-of-the-art method for various​​​‌ types of query. This‌ highlights that hybridization can‌​‌ be effectively modeled using​​ cosine similarity, without the​​​‌ need for training. This‌ work was published in‌​‌ 11.

8.1.6 SCALED​​ : Surrogate-gradient for Codec-Aware​​​‌ Learning of Downsampling in‌ ABR Streaming

Participants: Esteban‌​‌ Pesnel, Aline Roumy​​, Thomas Maugey.​​​‌

The rapid growth in‌ video consumption has intro-‌​‌ duced significant challenges to​​ modern streaming architectures. Over-the-Top​​​‌ (OTT) video delivery now‌ predominantly relies on Adaptive‌​‌ Bitrate (ABR) streaming, which​​ dynamically adjusts bitrate and​​​‌ resolution based on client-side‌ constraints such as display‌​‌ capabilities and network bandwidth.​​ This pipeline typically involves​​​‌ downsampling the original high-resolution‌ content, encoding and transmitting‌​‌ it, followed by decoding​​ and upsampling on the​​​‌ client side. Traditionally, these‌ processing stages have been‌​‌ optimized in isolation, leading​​ to suboptimal end-to-end rate-distortion​​​‌ (R-D) performance. The advent‌ of deep learning has‌​‌ spurred interest in jointly​​ optimizing the ABR pipeline​​​‌ using learned resampling methods.‌ However, training such systems‌​‌ end-to-end remains challenging due​​ to the non-differentiable nature​​​‌ of standard video codecs,‌ which obstructs gradient-based optimization.‌​‌ Recent works have addressed​​ this issue using dif-​​​‌ ferentiable proxy models, based‌ either on deep neural‌​‌ networks or hybrid coding​​ schemes with differentiable components​​​‌ such as soft quantization,‌ to approximate the codec‌​‌ behavior. While differentiable proxy​​ codecs have enabled progress​​​‌ in compression-aware learning, they‌ remain approximations that may‌​‌ not fully capture the​​ behavior of standard, non-differentiable​​​‌ codecs. To our knowledge,‌ there is no prior‌​‌ evidence demonstrating the inefficiencies​​ of using standard codecs​​​‌ during training. In this‌ work, we introduce a‌​‌ novel framework that enables​​ end-to- end training with​​​‌ real, non-differentiable codecs by‌ leveraging data-driven surrogate gradients‌​‌ derived from actual compression​​ errors. It facilitates the​​​‌ alignment between training objectives‌ and deployment performance. Experimental‌​‌ results show a 5.19improvement​​​‌ in BD-BR (PSNR) compared​ to codec-agnostic training approaches,​‌ consistently across the entire​​ rate-distortion convex hull spanning​​​‌ multiple downsampling ratios. This​ work was published in​‌ 15.

8.1.7 OSLO-IC:​​ On-the-Sphere Learned Omnidirectional Image​​​‌ Compression with Attention Modules​ and Spatial Context

Participants:​‌ Thomas Maugey.

Developing​​ effective 360-degree (spherical) image​​​‌ compression techniques is crucial​ for technologies like virtual​‌ reality and automated driving.​​ This work advances the​​​‌ state-of-the-art in on-the-sphere learning​ (OSLO) for omnidirectional image​‌ compression framework by proposing​​ spherical attention modules, residual​​​‌ blocks, and a spatial​ autoregressive context model. These​‌ improvements achieve a 23.1​​$\%$ bit rate reduction​​​‌ in terms of WS-PSNR​ BD rate. Additionally, we​‌ introduce a spherical transposed​​ convolution operator for upsampling,​​​‌ which reduces trainable parameters​ by a factor of​‌ four compared to the​​ pixel shuffling used in​​​‌ the OSLO framework, while​ main- taining similar compression​‌ performance. Therefore, in total,​​ our proposed method offers​​​‌ significant rate savings with​ a smaller architecture and​‌ can be applied to​​ any spherical convolutional application.​​​‌ This work was published​ in 18.

8.2.1 Semantic​​​‌ compression of images at​ extremely low bitrate

Participants:​‌ Tom Bordin, Thomas​​ Maugey.

We propose​​​‌ a framework for semantic​ image compression targeting ultra-low​‌ bitrates ($\sim$0.001​​ bpp). The semantic content​​​‌ of an image is​ transmitted through its representation​‌ in the CLIP embedding​​ space. Although embeddings lack​​​‌ positional information, semantic features​ provide strong priors that​‌ can be modeled with​​ attention layers (instead of​​​‌ color map as introduced​ in previous work). We​‌ leverage these priors to​​ transmit only residual positional​​​‌ data as attention maps,​ thereby correcting the spatial​‌ arrangement of objects in​​ the scene. Our method​​​‌ is evaluated using both​ standard objective metrics and​‌ subjective human assessments, demonstrating​​ state-of-the-art performance in both​​​‌ aspects. This work is​ currently under review.

However,​‌ in applications targeting extremely​​ low bitrates (0.01 bpp),​​​‌ where the reconstruction distortion​ can be severe, it​‌ makes sense to prioritize​​ parts of the image​​​‌ that are more relevant​ than others. In a​‌ second work, we propose​​ a semantic compression framework​​​‌ that integrates user or​ application preferences to compress​‌ image parts based on​​ their semantic representation. We​​​‌ design a guide for​ trained diffusion models that​‌ takes into account the​​ preferences for describing objects​​​‌ with varying accuracies. We​ show that we are​‌ able to preserve the​​ selected objects while also​​​‌ preserving the semantic and​ global aspect of the​‌ image without any retraining​​ or fine-tuning. This work​​​‌ is currently under review.​

8.2.2 Compressing image encoders​‌ via latent distillation

Participants:​​ Caroline Mazini-Rodrigues, Nicolas​​​‌ Keriven, Thomas Maugey​.

Deep learning models​‌ for image compression often​​ face practical limitations in​​​‌ hardware-constrained applications. Although these​ models achieve high-quality reconstructions,​‌ they are typically complex,​​ heavyweight, and require substantial​​​‌ training data and computational​ resources. We propose a​‌ methodology to partially compress​​ these networks by reducing​​​‌ the size of their​ encoders. Our approach uses​‌ a simplified knowledge distillation​​ strategy to approximate the​​ latent space of the​​​‌ original models with less‌ data and shorter training,‌​‌ yielding lightweight encoders from​​ heavy-weight ones. We evaluate​​​‌ the resulting lightweight encoders‌ across two different architectures‌​‌ on the image compression​​ task. Experiments show that​​​‌ our method preserves recon-‌ struction quality and statistical‌​‌ fidelity better than training​​ lightweight encoders with the​​​‌ original loss, making it‌ practical for resource-limited environments.‌​‌ This work is currently​​ under review 28.​​​‌

8.2.3 Energy-aware images via‌ pixel value reduction: the‌​‌ impact of compression on​​ attenuation maps

Participants: Emmanuel​​​‌ Sampaio, Thomas Maugey‌.

Video consumption accounts‌​‌ for a significant share​​ of global energy use,​​​‌ with end devices responsible‌ for most of it.‌​‌ On end devices, display​​ technology plays an important​​​‌ role in energy consumption.‌ Interestingly, OLED technology allows‌​‌ power to be adapted​​ via pixel-intensity manipulation. In​​​‌ this context, Pixel Value‌ Reduction (PVR) has shown‌​‌ promising results for lowering​​ display power by generating​​​‌ attenua- tion maps that‌ adapt image luminance. However,‌​‌ the use of this​​ technology in streaming services​​​‌ has not been fully‌ studied. In this work,‌​‌ we analyze the effect​​ of attenuation-map compression on​​​‌ perceptual quality, bitrate overhead,‌ and end-device energy consumption.‌​‌ Using a pixel-value- reduction​​ model, we generate attenuation​​​‌ maps for target power-reduction‌ levels (10$\%$,‌​‌ 20$\%$, and​​ 40$\%$) and​​​‌ encode them with the‌ HEVC video codec at‌​‌ various quantization- parameter (QP)​​ values (i.e., codec QP).​​​‌ Experiments on 4K content‌ with real OLED power‌​‌ measurements show that compressed​​ attenuation maps maintain high​​​‌ fidelity to the originals,‌ achieving different levels of‌​‌ power reduction with negligible​​ quality loss. Moreover, the​​​‌ results indicate that proper‌ alignment between content and‌​‌ map quantization pa- rameters​​ is critical for reducing​​​‌ bitrate overhead. These findings‌ highlight the feasibility of‌​‌ transmitting compressed attenuation maps​​ to minimize display's energy​​​‌ consumption. This work is‌ currently under review.

8.2.4‌​‌ Experimental analysis of the​​ impact of multi-threading on​​​‌ video encoding energy consumption‌

Participants: Robin Richard,‌​‌ Thomas Maugey.

Modern​​ CPUs are equipped with​​​‌ more and more cores,‌ raising the question of‌​‌ how parallelism leads to​​ better energy efficiency, especially​​​‌ in intensive tasks like‌ video encoding. This work‌​‌ investigates how video encoding​​ using multiple threads leads​​​‌ to better usage of‌ available cores, and if‌​‌ it actually improves energy​​ efficiency. Based on real​​​‌ video transcoding energy measurements‌ on a server, we‌​‌ test classical energy models​​ in a multi-threaded context.​​​‌ On the one hand,‌ we observe that the‌​‌ energy consumed during encoding​​ is indeed decreasing with​​​‌ the number of cores‌ used during the task.‌​‌ On the other hand,​​ we also observe that​​​‌ this number of used‌ cores is not always‌​‌ linked to the number​​ of threads that are​​​‌ given in parameter to‌ the encoder. Hence, this‌​‌ study enables to state​​ that the energy savings​​​‌ due to multi-threading is‌ likely for small number‌​‌ of threads, but less​​ achievable when the number​​​‌ of threads becomes too‌ high. This work is‌​‌ currently under review.

8.2.5​​​‌ Efficiency vs sufficiency for​ video streaming systems

Participants:​‌ Thomas Maugey, Anne-Cécile​​ Orgerie, Robin Richard​​​‌.

To reduce the​ ecological impact of a​‌ technology, scientists often focus​​ on energy efficiency issues,​​​‌ ignoring the complex rebound​ effects generated by efficiency.​‌ We focus on the​​ video transmission technology, and​​​‌ discuss the urgent need​ to be able to​‌ set limits in order​​ to target absolute sustainability​​​‌ and sufficiency. We show​ that these limits can​‌ provoke opposition or circumvention,​​ illustrating the difficulty of​​​‌ the task. We conclude​ that the question of​‌ limits must be considered​​ as a research problem​​​‌ in its own right,​ and that it is​‌ intrinsically multidisciplinary. This work​​ has been presented in​​​‌ 22.

8.2.6 Video​ streaming: how do the​‌ socio-economical models shape our​​ research questions?

Participants: Natacha​​​‌ Lapeyroux, Thomas Maugey​, Anne-Cécile Orgerie.​‌

According to a various​​ number of studies, the​​​‌ environmental and social impacts​ of video streaming is​‌ huge and growing. Today,​​ the work of researchers​​​‌ in the field of​ image processing only accelerates​‌ this explosion by contributing​​ to the emergence of​​​‌ new technologies. At best,​ researchers are simply trying​‌ to improve the efficiency​​ of streaming systems, which,​​​‌ due to the rebound​ effects, also contributes to​‌ “accelerating the acceleration”. In​​ this talk, we give​​​‌ an overview of the​ socio-economical models ruling most​‌ of the video streaming​​ platforms, and we show​​​‌ that the research questions​ tackled nowadays are directly​‌ shaped from these models.​​ We also show that​​​‌ these models irremediably lead​ to bigger videos and​‌ more videos. Tackling the​​ reduction of video streaming​​​‌ impacts will only be​ possible by questioning these​‌ models

8.3.1 GS-Morph:​​​‌ Dynamic Novel View Synthesis​ via UDF-ARAP Gaussian Splat​‌ Morphing,

Participants: Stephane Belemkoabga​​, Christine Guillemot,​​​‌ Thomas Maugey.

Monocular​ view synthesis in dynamic​‌ scenes remains a fundamental​​ challenge in vision and​​​‌ graphics, particularly for applications​ like augmented reality, virtual​‌ production, and free-viewpoint video.​​ Recovering accurate 3D geometry​​​‌ and realistic rendering from​ a single RGB-D stream​‌ is highly ill-posed due​​ to partial, noisy, and​​​‌ temporally inconsistent observations under​ non-rigid motion. Recent methods,​‌ such as dynamic NeRFs​​ and 4D Gaussian Splatting,​​​‌ attempt to jointly optimize​ motion and geometry. While​‌ effective near training trajectories,​​ these entangled designs often​​​‌ struggle to generalize across​ novel views and times.​‌ We introduce a new​​ framework that explicitly decouples​​​‌ geometry reconstruction and motion​ estimation to improve robustness​‌ and generalization. Given a​​ monocular RGB-D sequence with​​​‌ known poses, we first​ extract per-frame point clouds​‌ and estimate frame-to-frame deformation​​ fields using Unsigned Distance​​​‌ Field (UDF) registration with​ ARAP regularization. These are​‌ used to segment the​​ sequence into motion- coherent​​​‌ Groups of Pictures (GoPs).​ Each GoP undergoes alternating​‌ fusion and deformation propagation​​ to yield a consistent​​​‌ local geometry and dense​ deformation field. GoPs are​‌ then hierarchically merged into​​ a global scene model​​​‌ with a unified deformation​ field. A spatio-temporal 3D​‌ Gaussian Splatting representation is​​ initialized from this model​​ and further refined with​​​‌ photometric and geometric losses.‌ To evaluate generalization, we‌​‌ introduce a two-level protocol:​​ Level 1 tests novel​​​‌ views along the training‌ path, while Level 2‌​‌ tests novel views at​​ unseen times or poses.​​​‌ We also release a‌ new RGB-D dataset for‌​‌ monocular dynamic scene recon-​​ struction. Our method sets​​​‌ a new state-of-the-art, outperforming‌ prior work in both‌​‌ synthesis quality and deformation​​ accuracy. This work was​​​‌ published in 16.‌

8.3.2 CAFe-GS: Compactness-Aware Frequency-Guided‌​‌ Densification for 3D Gaussian​​ Splatting

Participants: Christine Guillemot​​​‌, Leo-Paul Huar.‌

3D Gaussian Splatting (3DGS)‌​‌ represents scenes using Gaus-​​ sian primitives and enables​​​‌ real-time novel view synthesis.‌ Adaptive Den- sity Control‌​‌ (ADC), a key part​​ of the pipeline, governs​​​‌ when to den- sify‌ these primitives to balance‌​‌ reconstruction quality and efficiency.​​ In the original 3DGS​​​‌ pipeline, densification is triggered‌ by a thresholded positional-gradient‌​‌ criterion. However, this criterion​​ frequently selects already well-covered​​​‌ regions, leading to redundant‌ primitives and pro- viding‌​‌ weak control over the​​ balance between reconstruction quality​​​‌ and compactness (i.e., fidelity‌ versus primitive count). In‌​‌ CAFe-GS, we pro- pose​​ a new densification criterion​​​‌ based on a per-Gaussian‌ score obtained by mapping‌​‌ per-pixel rendering errors back​​ to the contributing primi-​​​‌ tives, using their effective-opacity‌ under front-to-back alpha composit-‌​‌ ing as weights. The​​ score is then modulated​​​‌ by frequency guidance derived‌ from Laplacian-of-Gaussian responses, promoting‌​‌ detail-rich, high- frequency areas​​ in contrast to smooth​​​‌ or already well-reconstructed re-‌ gions. This criterion drives‌​‌ densification through standard cloning​​ and splitting operations. CAFe-GS​​​‌ provides a clearer, single-parameter‌ han- dle on the‌​‌ quality–compactness balance. Experiments on​​ standard benchmarks show that​​​‌ CAFe-GS achieves comparable PSNR‌ using 2 to 4‌​‌ times fewer Gaussians at​​ matched quality, and up​​​‌ to 12 to 15‌ times fewer Gaussians at‌​‌ a controlled PSNR trade-off.​​

8.3.3 Extended-Depth Multispectral Fluorescence​​​‌ Microscopy with Co-Designed Meta-optics‌ and Reconstruction

Participants: Ipek‌​‌ Anil Atalay Appak,​​ Christine Guillemot.

Fluorescence​​​‌ microscopy can deliver high-resolution‌ spatial details; however, it‌​‌ suffers from shallow depth​​ of field and chromatic​​​‌ aberrations. The impact is‌ greatest for thick specimens‌​‌ and for multispectral data​​ that must stay aligned​​​‌ across depth. We have‌ designed MANTIS (Multispectral All-Depth‌​‌ meta-opTics Imaging System), a​​ co-designed optical–computational platform that​​​‌ achieves extended depth of‌ field from a single‌​‌ acquisition per field of​​ view without axial scanning.​​​‌ A learned meta-optic and‌ a physics-guided reconstruction are‌​‌ trained end-to-end so that​​ depth and wavelength-dependent blur​​​‌ is encoded in a‌ recoverable form and decoded.‌​‌ We target extended depth​​ ranges reaching up to​​​‌ 75 micrometer. The reconstructions‌ show weak depth dependence‌​‌ and low cross-spectral variance.​​ In simulation at 50​​​‌ micrometer depth of field,‌ mean peak signal-to-noise ratio‌​‌ and structural similarity reach​​ 23.5 dB and 0.70,​​​‌ averaged over depths and‌ channels. We have validated‌​‌ experimentally the designed system​​ by fabricating the learned​​​‌ meta-optic, measuring the point‌ spread functions across the‌​‌ target depths and wavelengths,​​ and reconstructing three-dimensional fluorescence​​​‌ samples. The experimental reconstructions‌ maintain contrast and lateral‌​‌ sharpness across depth, exhibiting​​​‌ modest per-channel variation in​ PSNR and SSIM, with​‌ trends that match the​​ simulation and are consistent​​​‌ with low chromatic aberration​ and extended depth of​‌ field.

8.4.1 MUPET: Maximum A​ Posteriori Training of Diffusion​‌ Models for Image Restoration​​

Participants: Christine Guillemot,​​​‌ Samuel Willingham.

Inverse​ problems involve reconstructing clean​‌ images from degraded observations.​​ Maximum a Posteriori (MAP)​​​‌ estimation reconstructs the most​ probable source image from​‌ noisy measurements. When combined​​ with Plug-and-Play (PnP) priors​​​‌ defined by an image​ denoising algorithm, MAP estimation​‌ yields high-quality reconstructions. In​​ contrast, Diffusion Models (DMs)​​​‌ address inverse problems by​ sampling from the posterior​‌ distribution using score functions​​ trained on images perturbed​​​‌ by Gaussian noise. Prior​ work reformulated diffusion sampling​‌ as Deep Equilibrium (DEQ)​​ models but did not​​​‌ fine-tune DMs for inverse​ problems. We have proposed​‌ MaximUm a PostEriori Training​​ (MUPET), a framework that​​​‌ leverages PnP gradient descent​ to enable DEQ fine-tuning​‌ of DMs on inverse​​ problems 19. By​​​‌ refining a generative prior​ at the fixed-point of​‌ MAP estimation, MUPET enhances​​ image restoration via posterior​​​‌ sampling while maintaining quality​ when sampling from the​‌ prior.

8.4.2 Taxonomy of​​ reduction matrices for Graph​​​‌ Coarsening

Participants: Antonin Joly​, Nicolas Keriven,​‌ Aline Roumy.

Graph​​ coarsening aims to diminish​​​‌ the size of a​ graph to lighten its​‌ memory footprint, and has​​ numerous applications in graph​​​‌ signal processing and machine​ learning. It is usually​‌ defined using a reduction​​ matrix and a lifting​​​‌ matrix, which, respectively, allows​ to project a graph​‌ signal from the original​​ graph to the coarsened​​​‌ one and back. This​ results in a loss​‌ of information measured by​​ the so-called Restricted Spectral​​​‌ Approximation (RSA). Most coarsening​ frameworks impose a fixed​‌ relationship between the reduction​​ and lifting matrices, generally​​​‌ as pseudo-inverses of each​ other, and seek to​‌ define a coarsening that​​ minimizes the RSA. In​​​‌ 13, we remark​ that the roles of​‌ these two matrices are​​ not entirely symmetric: indeed,​​​‌ putting constraints on the​ lifting matrix alone ensures​‌ the existence of important​​ objects such as the​​​‌ coarsened graph's adjacency matrix​ or Laplacian. In light​‌ of this, in this​​ paper, we introduce a​​​‌ more general notion of​ reduction matrix, that is​‌ not necessarily the pseudo-inverse​​ of the lifting matrix.​​​‌ We establish a taxonomy​ of “admissible” families of​‌ reduction matrices, discuss the​​ different properties that they​​​‌ must satisfy and whether​ they admit a closed-form​‌ description or not. We​​ show that, for a​​​‌ fixed coarsening represented by​ a fixed lifting matrix,​‌ the RSA can be​​ further reduced simply by​​​‌ modifying the reduction matrix.​ We explore different examples,​‌ including some based on​​ a constrained optimization process​​​‌ of the RSA. Since​ this criterion has also​‌ been linked to the​​ performance of Graph Neural​​​‌ Networks, we also illustrate​ the impact of this​‌ choices on different node​​ classification tasks on coarsened​​​‌ graphs. This work was​ published at the NeurIPS​‌ conference.

8.4.3 Node Regression​​ on Latent Position Random​​ Graphs via Local Averaging​​​‌

Participants: Nicolas Keriven.‌

Node regression consists in‌​‌ predicting the value of​​ a graph label at​​​‌ a node, given observations‌ at the other nodes.‌​‌ To gain some insight​​ into the performance of​​​‌ various estimators for this‌ task, in 7 we‌​‌ perform a theoretical study​​ in a context where​​​‌ the graph is random.‌ Specifically, we assume that‌​‌ the graph is generated​​ by a Latent Position​​​‌ Model, where each node‌ of the graph has‌​‌ a latent position, and​​ the probability that two​​​‌ nodes are connected depend‌ on the distance between‌​‌ the latent positions of​​ the two nodes. In​​​‌ this context, we begin‌ by studying the simplest‌​‌ possible estimator for graph​​ regression, which consists in​​​‌ averaging the value of‌ the label at all‌​‌ neighboring nodes. We show​​ that in Latent Position​​​‌ Models this estimator tends‌ to a Nadaraya Watson‌​‌ estimator in the latent​​ space, and that its​​​‌ rate of convergence is‌ in fact the same.‌​‌ One issue with this​​ standard estimator is that​​​‌ it averages over a‌ region consisting of all‌​‌ neighbors of a node,​​ and that depending on​​​‌ the graph model this‌ may be too much‌​‌ or too little. An​​ alternative consists in first​​​‌ estimating the true distances‌ between the latent positions,‌​‌ then injecting these estimated​​ distances into a classical​​​‌ Nadaraya Watson estimator. This‌ enables averaging in regions‌​‌ either smaller or larger​​ than the typical graph​​​‌ neighborhood. We show that‌ this method can achieve‌​‌ standard nonparametric rates in​​ certain instances even when​​​‌ the graph neighborhood is‌ too large or too‌​‌ small. This work was​​ published in the Journal​​​‌ of Machine Learning Research‌ (JMLR).

8.4.4 Backward Oversmoothing:‌​‌ why is it hard​​ to train deep Graph​​​‌ Neural Networks?

Participants: Nicolas‌ Keriven.

Oversmoothing has‌​‌ long been identified as​​ a major limitation of​​​‌ Graph Neural Networks (GNNs):‌ input node features are‌​‌ smoothed at each layer​​ and converge to a​​​‌ non-informative representation, if the‌ weights of the GNN‌​‌ are sufficiently bounded. This​​ assumption is crucial: if,​​​‌ on the contrary, the‌ weights are sufficiently large,‌​‌ then oversmoothing may not​​ happen. Theoretically, GNN could​​​‌ thus learn to not‌ oversmooth. However it does‌​‌ not really happen in​​ practice, which prompts us​​​‌ to examine oversmoothing from‌ an optimization point of‌​‌ view. In the preprint​​ 27, we analyze​​​‌ backward oversmoothing, that is,‌ the notion that backpropagated‌​‌ errors used to compute​​ gradients are also subject​​​‌ to oversmoothing from output‌ to input. With non-linear‌​‌ activation functions, we outline​​ the key role of​​​‌ the interaction between forward‌ and backward smoothing. Moreover,‌​‌ we show that, due​​ to backward oversmoothing, GNNs​​​‌ provably exhibit many spurious‌ stationary points: as soon‌​‌ as the last layer​​ is trained, the whole​​​‌ GNN is at a‌ stationary point. As a‌​‌ result, we can exhibit​​ regions where gradients are​​​‌ near-zero while the loss‌ stays high. The proof‌​‌ relies on the fact​​ that, unlike forward oversmoothing,​​​‌ backward errors are subjected‌ to a linear oversmoothing‌​‌ even in the presence​​​‌ of non-linear activation function,​ such that the average​‌ of the output error​​ plays a key role.​​​‌ Additionally, we show that​ this phenomenon is specific​‌ to deep GNNs, and​​ exhibit counter-example Multi-Layer Perceptron.​​​‌ This paper is a​ step toward a more​‌ complete comprehension of the​​ optimization landscape specific to​​​‌ GNNs.

9.1.1 CIFRE contract​​​‌ with TyndallFx on Radiance​ fields representation for dynamic​‌ scene reconstruction

Participants: Christine​​ Guillemot [contact], Stephane​​​‌ Belemkoabga, Thomas Maugey​.

• Title  : Radiance​‌ fields representation for dynamic​​ scene reconstruction
• Partners :​​​‌ TyndallFx (R. Mallart), Inria-Rennes.​
• Funding : TyndallFx, ANRT.​‌
• Period : Oct-2023-June. 2025​​

The goal of this​​​‌ project is to design​ novel methods for modeling​‌ and compact representation of​​ radiance fields for scene​​​‌ reconstruction and view synthesis.​ The problems that are​‌ addressed are those of​​ fast and efficient estimation​​​‌ of the camera pose​ parameters and of the​‌ 3D model of the​​ sceen based on Gaussian​​​‌ splatting, and as as​ the one of tracking​‌ and modeling the deformation​​ of the model due​​​‌ to the global camera​ motion and to the​‌ motion of the different​​ objects in the scene.​​​‌

9.1.2 CIFRE contract with​ MediaKind on Learned video​‌ downscaling for end-to-end Rate-Distortion​​ optimization of video streaming​​​‌ system

Participants: Thomas Maugey​ [contact], Esteban Pesnel​‌, Aline Roumy.​​

• Title  : Learned video​​​‌ downscaling for end-to-end Rate-Distortion​ optimization of video streaming​‌ system
• Partners : MediaKind,​​ Inria-Rennes.
• Funding : MediaKind,​​​‌ ANRT.
• Period : November​ 2023-October 2026.

This CIFRE​‌ contract aims to optimize​​ a streaming solution by​​​‌ addressing constraints related to​ distribution, standards, and deployment.​‌ The focus is on​​ developing downscaling techniques that​​​‌ enhance the end-to-end streaming​ process, considering bitrate-distortion optimization.​‌ While the upscaling filter​​ on client devices is​​​‌ fixed due to standardization,​ encoding and downscaling on​‌ the server side remain​​ flexible, offering an opportunity​​​‌ for improvement within the​ streaming pipeline.

9.1.3 CIFRE​‌ contract with InterDigital on​​ Hybrid conventional and deep​​​‌ learning-based video coding

Participants:​ Aline Roumy [contact],​‌ Antoine Monier.

• Title​​  : Hybrid conventional and​​​‌ deep learning-based video coding​
• Partners : InterDigital, Inria-Rennes.​‌
• Funding : InterDigital, ANRT.​​
• Period : Jan. 2025-Dec.​​​‌ 2028.

This CIFRE contract​ aims to improve conventional​‌ video codecs in terms​​ of compression efficiency with​​​‌ the help of deep-learning​ and machine-learning based coding​‌ tools. The goal is​​ to investigate the usage​​​‌ of deep-learning solutions for​ enhancing core video coding​‌ modules such as transform​​ and residual (transform coefficients)​​​‌ coding, in-loop filtering, prediction.​ These new solutions should​‌ complement or replace existing​​ coding tools or modes,​​​‌ such as the ones​ implemented in the VVC​‌ standard, or in the​​ exploratory video coding model​​​‌ developed by the JVET​ standardization group named "Enhanced​‌ coding model" (ECM).

9.1.4​​ CIFRE contract with InterDigital​​​‌ on End-to-end energy-constrained video​ content delivery

Participants: Thomas​‌ Maugey [contact], Emmanuel​​ Sampaio.

• Title  :​​​‌ End-to-end energy-constrained video content​ delivery
• Partners : InterDigital,​‌ Inria-Rennes.
• Funding : InterDigital,​​ ANRT.
• Period : Jan.​​ 2025-Dec. 2028.

The goal​​​‌ is to investigate new‌ algorithms and video delivery‌​‌ frameworks to reduce the​​ energy consumption footprint (and​​​‌ then the carbon footprint)‌ of video content delivery.‌​‌ To reach this ambitious​​ goal, several levers or​​​‌ strategies can be activated:‌

• Content pre-processing for reducing‌​‌ the encoding / transmission​​ / decoding / rendering​​​‌ energy footprint. Assuming that‌ the content is modified‌​‌ at the server side,​​ this raises some important​​​‌ concerns: can we maintain‌ the Quality of Experience?‌​‌ Can we guarantee an​​ acceptance level? Do we​​​‌ need to provide side-information‌ for making the process‌​‌ more efficient? If yes,​​ is this overhead relevant​​​‌ for a commercial and‌ viable operational deployment?
• Content‌​‌ post-processing for reducing the​​ rendering energy footprint. Modifying​​​‌ the content at the‌ client side raises the‌​‌ concern of the computational​​ cost. A balance between​​​‌ energy gain and energy‌ required to perform the‌​‌ post-processing operation has to​​ be carefully considered.
• The​​​‌ delivery and consumption of‌ video content are performed‌​‌ thanks to video streaming​​ services. One of the​​​‌ key ingredients of such‌ services relies on adaptive‌​‌ bitrate techniques aiming to​​ deliver the highest QoE​​​‌ to the users given‌ a bit rate constraint.‌​‌ We may want to​​ go further by adding​​​‌ a new ingredient to‌ the recipe, i.e., the‌​‌ energy consumption of such​​ services. By considering the​​​‌ bit rate, the quality‌ of experience and the‌​‌ energy footprint of the​​ video, new energy-aware video​​​‌ streaming services could be‌ envisioned

10.1.1 Horizon Europe

Participants:​​​‌ Nicolas Keriven [PI],‌ Hugo Jaquard, Adarsh‌​‌ Jamadandi.

ERC Starting​​ Grant MALAGA: Reinventing the​​​‌ Theory of Machine Learning‌ on Large Graphs

• Period:‌​‌ 2025 - 2030

In​​ many scientific domains, graphs​​​‌ are the objects of‌ choice to represent structured‌​‌ data: from molecules to​​ social networks, power grids,​​​‌ the internet, and so‌ on. The exploitation of‌​‌ graph data represents a​​ major scientific and industrial​​​‌ challenge. Graph Machine Learning‌ (Graph ML) is thus‌​‌ a fast-growing field, with​​ so-called Graph Neural Networks​​​‌ (GNN) at the forefront.‌

However, in sharp contrast‌​‌ with traditional ML, the​​ field of GML has​​​‌ somewhat jumped from early‌ methods to deep learning,‌​‌ without the decades-long development​​ of well-established notions to​​​‌ compare, analyze and improve‌ algorithms. As a result,‌​‌ GNNs have limitations, both​​ practical and theoretical, and​​​‌ it is not clear‌ how to address them.‌​‌ Practical results may vary​​ wildly depending on the​​​‌ architecture and datasets, with‌ no guidelines on how‌​‌ to design reliable GNNs​​ in each case. Overall,​​​‌ these are the symptoms‌ of a major issue:‌​‌ Graph ML is somewhat​​ lacking fundamental theory. The​​​‌ ambition of project MALAGA‌ is to develop such‌​‌ a theory. Solving the​​ crucial limitations of the​​​‌ current theory is highly‌ challenging: fundamental mathematical tools‌​‌ in cannot analyze the​​ learning capabilities of Graph​​​‌ ML methods in a‌ unified way (e.g., graph‌​‌ nodes are not iid),​​ existing *statistical graph models*​​​‌ do not faithfully represent‌ the many characteristics of‌​‌ modern graph data (especially​​​‌ node features and their​ relationship with graph structure​‌ in homophilic and heterophilic​​ graphs), and computational complexity​​​‌ may become problematic on​ large graphs. MALAGA will​‌ develop a radically new​​ understanding of GML problems,​​​‌ and of the strengths​ and limitations of a​‌ large panel of algorithms.​​

10.1.2 H2020 projects

Participants:​​​‌ Christine Guillemot [contact],​ Anil Ipek Atalay Appak​‌, Soheib Takhtardeshir,​​ Samuel Willingham.

• Title:​​​‌ Plenoptima: Plenoptic Imaging
• Duration:​ From January 1, 2021​‌ to December 31, 2025​​
• Partners:
  • INSTITUT NATIONAL DE​​​‌ RECHERCHE EN INFORMATIQUE ET​ AUTOMATIQUE (INRIA), France
  • MITTUNIVERSITETET​‌ (MIUN), Sweden
  • TECHNISCHE UNIVERSITAT​​ BERLIN (TUB), Germany
  • TAMPEREEN​​​‌ KORKEAKOULUSAATIO SR (TAMPERE UNIVERSITY),​ Finland
  • "INSTITUTE OF OPTICAL​‌ MATERIALS AND TECHNOLOGIES ""ACADEMICIAN​​ JORDAN MALINOWSKI"" - BULGARIAN​​​‌ ACADEMY OF SCIENCES" (IOMT),​ Bulgaria
• Inria contact: Christine​‌ Guillemot
• Coordinator: Tampere University​​ (Finland, Atanas Gotchev)

Plenoptic​​​‌ Imaging aims at studying​ the phenomena of light​‌ field formation, propagation, sensing​​ and perception along with​​​‌ the computational methods for​ extracting, processing and rendering​‌ the visual information.

The​​ PLENOPTIMA ultimate project goal​​​‌ is to establish new​ cross-sectorial, international, multi-university sustainable​‌ doctoral degree programmes in​​ the area of plenoptic​​​‌ imaging and to train​ the first fifteen future​‌ researchers and creative professionals​​ within these programmes for​​​‌ the benefit of a​ variety of application sectors.​‌ PLENOPTIMA develops a cross-disciplinary​​ approach to imaging, which​​​‌ includes the physics of​ light, new optical materials​‌ and sensing principles, signal​​ processing methods, new computing​​​‌ architectures, and vision science​ modelling. With this aim,​‌ PLENOPTIMA joints five of​​ strong research groups in​​​‌ nanophotonics, imaging and machine​ learning in Europe with​‌ twelve innovative companies, research​​ institutes and a pre-competitive​​​‌ business ecosystem developing and​ marketing plenoptic imaging devices​‌ and services.

PLENOPTIMA advances​​ the plenoptic imaging theory​​​‌ to set the foundations​ for developing future imaging​‌ systems that handle visual​​ information in fundamentally new​​​‌ ways, augmenting the human​ perceptual, creative, and cognitive​‌ capabilities. More specifically, it​​ develops 1) Full computational​​​‌ plenoptic imaging acquisition systems;​ 2) Pioneering models and​‌ methods for plenoptic data​​ processing, with a focus​​​‌ on dimensionality reduction, compression,​ and inverse problems; 3)​‌ Efficient rendering and interactive​​ visualization on immersive displays​​​‌ reproducing all physiological visual​ depth cues and enabling​‌ realistic interaction.

All ESRs​​ are registered in Joint/Double​​​‌ degree doctoral programmes at​ academic institutions in Bulgaria,​‌ Finland, France, Germany and​​ Sweden. The programmes will​​​‌ be made sustainable through​ a set of measures​‌ in accordance with the​​ Salzburg II Recommendations of​​​‌ the European University Association.​

10.2.1​‌ PEPR MoleculArXiv. Targeted project​​ 2: From Digital Data​​​‌ to Synthetic DNA

Participants:​ Aline Roumy [contact],​‌ Sara Al Sayyed,​​ Thomas Maugey.

• Partners:​​​‌ I3S, LabSTIC, IMT-Atlantique, Irisa/Inria​ (GenScale and Compact team),​‌ IPMC, Eurecom.
• Funding: France​​ 2030.
• Period: Sept. 2022​​​‌ - Feb. 2032.

The​ PEPR MoleculArXiv aims to​‌ develop future data storage​​ devices on molecular media,​​​‌ including DNA and artificial​ polymers. This involves not​‌ only parallelizing synthesis devices​​ but also discovering new​​​‌ molecules and information technologies​ to accelerate the synthesis​‌ of storage media, their​​ encoding and decoding, and​​ exploring various molecular supports.​​​‌

Within the targeted project‌ "From Digital Data to‌​‌ Synthetic DNA," the objective​​ is to make physical​​​‌ and logical storage efficient‌ through custom-designed codes tailored‌​‌ to the physicochemical constraints​​ of DNA writing and​​​‌ reading. This effort is‌ conducted in collaboration with‌​‌ partners from other targeted​​ projects, such as "Next-Generation​​​‌ DNA Synthesis" and "Synthetic‌ Digital Polymers."

Several key‌​‌ challenges are addressed, including​​ robustness to noise. Processes​​​‌ like synthesis, sequencing, storage,‌ or manipulation of DNA‌​‌ can introduce errors that​​ threaten the integrity of​​​‌ the stored data. These‌ errors are non-classical compared‌​‌ to those encountered in​​ wired and wireless communication​​​‌ channels and require specific‌ handling. This issue is‌​‌ approached from the perspectives​​ of both compression and​​​‌ error-correcting codes.

Another critical‌ challenge is data access.‌​‌ A significant advantage of​​ storing information on DNA,​​​‌ apart from its durability,‌ is its extremely high‌​‌ density, enabling vast amounts​​ of data to be​​​‌ stored compactly. Due to‌ this high density, it‌​‌ is essential to facilitate​​ rapid access to the​​​‌ required data items. New‌ data representations are studied‌​‌ to enable fast random​​ access to the data​​​‌ relying merely on biological‌ and chemical processes.

10.2.2‌​‌ PEPR IA. Project SHARP​​ : Sharp Theoretical and​​​‌ Algorithmic Principles for frugal‌ ML

Participants: Nicolas Keriven‌​‌ [contact], Antonin Joly​​, Caroline Mazini-Rodrigues.​​​‌

• Partners: LIP, ENPC, IRISA,‌ INRIA, CEA, LAMSADE, ISIR.‌​‌
• Funding: France 2030.
• Period:​​ 2023 - 2028

SHARP​​​‌ will address the major‌ challenge of designing, analyzing‌​‌ and deploying a new​​ generation of intrinsically frugal​​​‌ models (neural or not)‌ able to achieve the‌​‌ versatility and performance of​​ today’s best models while​​​‌ requiring only a vanishing‌ fraction of the resources‌​‌ currently needed. This will​​ be achieved by the​​​‌ constitution of a strong‌ task force able to‌​‌ cover an integrated pipeline,​​ from theoretical foundations to​​​‌ flagship AI domains such‌ as computer vision and‌​‌ natural language processing. With​​ foundational advances towards stronger​​​‌ principles, smaller models, smaller‌ datasets, SHARP will allow‌​‌ tomorrow’s best AI systems​​ to run on yesterday’s​​​‌ devices, somewhat providing a‌ cure against obsolescence.

10.2.3‌​‌ ANR Young researcher grant:​​ MAssive multimedia DAta collection​​​‌ REpurposing (MADARE)

Participants: Thomas‌ Maugey [contact], Tom‌​‌ Bordin.

• Funding: ANR​​ (Agence Nationale de la​​​‌ Recherche)
• Period: pr. 2022‌ - Oct. 2025.

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 framework, proposed in​​ the MADARE project, will​​​‌ tear down this barrier,‌ by 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. Concretely, a​​​‌ data collection is only‌ encoded to a compact‌​‌ description that is used​​ to guarantee that the​​​‌ regenerated content is semantically‌ coherent with the initial‌​‌ one. In practice, it​​ opens several research directions:​​​‌ how to organise the‌ latent space (in which‌​‌ the coded descriptions lie)​​​‌ such that the information​ is efficiently and intelligibly​‌ represented? How to regenerate​​ a synthesized content from​​​‌ this compact description (based​ for example on guided​‌ diffusion algorithms)? Finally, how​​ to extend this idea​​​‌ to video? By revisiting​ the compression problem, the​‌ MADARE project aims gigantic​​ compression ratios enabling, among​​​‌ other benefits, to reduce​ the impact of exploding​‌ data creation on the​​ cloud servers’ energy consumption.​​​‌

10.2.4 Joint Project (Défi​ commun) Nisk.AI

Participants: Aline​‌ Roumy [contact], Thomas​​ Maugey, Antoine Monier​​​‌, Christine Guillemot,​ Emmanuel Victor Barbosa Sampaio​‌.

• Partners: Inria teams​​ (Compact, Combo, Taran), InterDigital.​​​‌
• Funding: Inria InterDigital.
• Period:​ Sept. 2022 - Feb.​‌ 2032.

Nisk.AI (2020-2026) is​​ a joint project with​​​‌ InterDigital on Sustainable Neural​ Network video coding. Indeed,​‌ video distribution faces two​​ major revolutions. The first​​​‌ one is due to​ the impact of AI​‌ technologies and in particular​​ deep learning. New ways​​​‌ to represent images and​ video have been proposed​‌ by the scientific community​​ and might impact how​​​‌ content is encoded, with​ very promising outputs in​‌ terms of coding efficiency​​ (e.g. the tradeoff between​​​‌ data-rate reduction and rendered​ perceived quality). The second​‌ revolution is the environmental​​ impact of media consumption,​​​‌ and more generally of​ ICT (Information and Communication​‌ Technologies), on the global​​ carbon footprint. This relates​​​‌ not only to the​ profusion of content and​‌ of its wide distribution,​​ but also to how​​​‌ this content is processed​ and consumed, including users’​‌ behavior. The first revolution​​ also has an impact​​​‌ on the second one​ due to the increased​‌ complexity of deep learning​​ architectures compared to conventional​​​‌ coding schemas. The objective​ of this project is​‌ to address those challenges​​ by proposing new deep-based​​​‌ video representation formats and​ coding schemes, taking into​‌ account efficiency, complexity and​​ sustainability. Both 2D and​​​‌ immersive video will be​ considered.

10.3.1 CominLabs Colearn project:​​ Coding for Learning

Participants:​​​‌ Aline Roumy [contact],​ Rémi Piau, Thomas​‌ Maugey.

• Partners: Inria-Rennes​​ (Compact team); LabSTICC, IMT​​​‌ Atlantique, (team Code and​ SI3); IETR, INSA Rennes​‌ (Syscom team).
• Funding: Labex​​ CominLabs.
• Period: Sept. 2021​​​‌ - Dec. 2026.
• contact:​ Aline Roumy

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.

10.3.2 CominLabs​​ VideoImpact project: Model the​​​‌ environmental cost of video​ delivery

Participants: Thomas Maugey​‌ [contact], Natacha Lapeyroux​​, Robin Richard.​​

• Partners: MAGELLAN (IRISA/Inria), VAADER​​​‌ at IETR/INSA, ARENES University‌ of Rennes, UCO Nantes,‌​‌ IMT Atlantique
• Funding: Labex​​ CominLabs.
• Period: Sept. 2025​​​‌ - Sep. 2027
• contact:‌ Thomas Maugey

Recent studies‌​‌ forecast a global warming​​ of 3.1°C in 2100​​​‌ if the GHG emissions‌ do not decrease. Hence,‌​‌ every part of our​​ society must urgently aim​​​‌ sobriety, including the digital‌ world, that is not‌​‌ intangible, contrary to popular​​ belief. Video consumption takes​​​‌ a significant part among‌ the emissions of the‌​‌ digital world and constitutes​​ a representative example of​​​‌ unbounded and energy-consuming digital‌ system. In that context,‌​‌ a crucial question to​​ tackle is how to​​​‌ set limits to the‌ deployment of a digital‌​‌ system, and for example​​ to video delivery systems?​​​‌ This question is, by‌ nature, lying at the‌​‌ crossroad of many fields​​ (including human and social​​​‌ sciences). Interestingly, many initiatives‌ have recently emerged at‌​‌ the regional level, e.g.,​​ the rapprochement between the​​​‌ GIS Marousin and video‌ processing scientists of INSA‌​‌ and IRISA, and set​​ interesting perspectives of wide​​​‌ collaborative user experiments. In‌ that context, the VideoImpact‌​‌ project proposes to answer​​ the following questions: In​​​‌ order to set a‌ sobriety policy, what should‌​‌ we limit in priority?​​ the number of hours​​​‌ spent by a user‌ watching videos? The TV‌​‌ screen size? The video​​ resolutions? The deployment of​​​‌ more efficient digital infrastructure?‌ The VideoImpact project aims‌​‌ at developing i) an​​ environmental footprint model for​​​‌ the video delivery chain‌ to identify the clear‌​‌ levers to sobriety, ii)​​ a solid network of​​​‌ industrial and academic partners‌ of the Rennes' neighborhood‌​‌ around the goal of​​ reducing the environmental impact​​​‌ of video consumption and‌ iii) to launch a‌​‌ concrete experimentation in collaboration​​ with Human and Social​​​‌ scientists. The conclusions will‌ be used in the‌​‌ context of further collaborations​​ with Human and Social​​​‌ Scientists to set real‌ user experiments to assess‌​‌ the feasibility and acceptance​​ of such levers.

10.3.3​​​‌ ARED VideoLimit project

Participants:‌ Thomas Maugey [contact],‌​‌ Robin Richard.

• Partners:​​ MAGELLAN (IRISA/Inria)
• Funding: Labex​​​‌ CominLabs.
• Period: Sept. 2025‌ - Sep. 2028

In‌​‌ line with the project​​ Cominlabs VideoImpact, the project​​​‌ Vlimit will specifically focus‌ on the modeling of‌​‌ the energetic expense of​​ the video transmission chain.​​​‌ More specifically the thesis‌ funded by the project‌​‌ will focus on:

• model​​ the energy spent over​​​‌ the whole video processing‌ chain during different delivery‌​‌ scenarios, based on the​​ state-of-the art analysis and​​​‌ experimental measurement campain.
• identify‌ the high-energetic parts in‌​‌ this pipeline and some​​ related levers that could​​​‌ be put in place‌ to reduce their costs,‌​‌ based on a simulation​​ tool for a «​​​‌ what-if » analysis.
• discuss‌ with the Human and‌​‌ Social Siences researchers for​​ setting the foundations of​​​‌ experimentations and inter-discplinary research‌ directions, based on regular‌​‌ meetings and workshops with​​ the active regional community​​​‌

11.1.1 Scientific‌​‌ events: organisation

11.1.2​​ Journal

11.1.3 Invited talks​​​‌

• Thomas Maugey gave a​ talk at L2S, Paris​‌ Saclay on "semantic compression:​​ exploring ultra low bitrate"​​​‌ (January)
• Thomas Maugey gave​ a talk at the​‌ GdR meeting on "Sustainaibility​​ and carbon footprint of​​​‌ the video transmission chain"​ on “Reducing environmental impact:​‌ from global modeling to​​ behavioral change” (March)
• Thomas​​​‌ Maugey gave a talk​ at the VAADER semainar​‌ (IINSA IETR), on “Reducing​​ environmental impact: from global​​​‌ modeling to behavioral change”​ (May)
• Thomas Maugey gave​‌ a talk at the​​ Inria-InterDigital NEMO workshop on​​​‌ "semantic compression: exploring ultra​ low bitrate" (November)
• Aline​‌ Roumy gave a tutorial​​ on “Information theory for​​​‌ image and video compression:​ fundamental results and recent​‌ challenges" MoleculArXiv Autumn School​​ on DNA Data Storage,​​​‌ Nov. 2025.
• Aline Roumy​ gave a talk at​‌ the Inria-InterDigital NEMO workshop​​ on "Image compression at​​​‌ JPEG: JPEG AI and​ JPEG DNA" (Nov. 2025)​‌

11.1.4 Leadership within the​​ scientific community

• Thomas Maugey​​​‌ is Vice-Chair of the​ EURASIP Technical Area Committee​‌ on Visual Information Processing​​
• Aline Roumy is a​​​‌ member of the IEEE​ Image, Video, and Multidimensional​‌ Signal Processing Technical Committee​​ (IVMSP TC).
• Aline Roumy​​​‌ is a member of​ the Executive board of​‌ the National Research group​​ in Image and Signal​​​‌ Processing (GRETSI).

11.1.5 Scientific​ expertise

• Christine Guillemot is​‌ member of the ERC​​ PE7 Advanced grant panel.​​​‌
• Christine Guillemot is member​ of the jury for​‌ the signal image vision​​ PhD prize of the​​​‌ Club EEA, GdR IASIS​ and GRETSI.
• Aline Roumy​‌ has been a member​​ of the jury for​​​‌ the recruitment of Inria​ Junior researcher (CRCN/ISFP) in​‌ Rennes, May 2025.
• Aline​​ Roumy served as a​​​‌ member of Board of​ Examiners (Comité de sélection)​‌ for an assistant professor​​ position (Maitres de Conférences)​​​‌ at Polytech Nantes University,​ May 2025.
• Aline Roumy​‌ has been a member​​ of the committee for​​​‌ the French Academy of​ Sciences/Inria Awards, June 2025.​‌
• Aline Roumy was a​​ reviewer for the evaluation​​ committee for the appointment​​​‌ of a professor, Telecom‌ Paris, Sept. 2025.
• Aline‌​‌ Roumy served as a​​ member of Board of​​​‌ Examiners (Comité de sélection)‌ for a Professor position‌​‌ (Professeur des Universités) at​​ CentraleSupélec, University, Oct. 2025.​​​‌

11.1.6 Research administration

• Christine‌ Guillemot is member of‌​‌ the ERC Cell of​​ the DPE (Direction des​​​‌ Programmes Européens) of Inria.‌
• Aline Roumy is a‌​‌ member of the research​​ commission and of the​​​‌ academic board of the‌ University of Rennes 2,‌​‌ as Inria representative
• Aline​​ Roumy is the co-director​​​‌ of the joint Inria/InterDigital‌ project (défi) Nisk.AI

11.2.1 Supervision​​​‌

• Thomas Maugey and Christine‌ Guillemot were co-supervising the‌​‌ PhD thesis of Stéphane​​ Belemkoabga in the context​​​‌ of the Cifre contract‌ with TyndallFX.
• Thomas Maugey‌​‌ and Aline Roumy are​​ co-supervising the PhD thesis​​​‌ of Esteban Pesnel in‌ the context of the‌​‌ Cifre contract with Mediakind.​​
• Thomas Maugey and Aline​​​‌ Roumy were co-supervising the‌ PhD thesis of Rémi‌​‌ Piau in the context​​ of the Cominlabs project​​​‌ CoLearn.
• Thomas Maugey and‌ Aline Roumy are co-supervising‌​‌ the PhD thesis of​​ Sara Al Sayyed in​​​‌ the context of the‌ PEPR project MoleculArxiv.
• Thomas‌​‌ Maugey is co-supervising the​​ PhD thesis of Emmanuel​​​‌ Sampaio in the context‌ of the Cifre contract‌​‌ with InterDigital.
• Thomas Maugey​​ was supervising the PhD​​​‌ thesis of Tom Bordin‌ in the context of‌​‌ the ANR project MADARE​​
• Thomas Maugey is co-supervising​​​‌ the PhD thesis of‌ Robin Richard in the‌​‌ context of the Bretagne​​ ARED contract.
• Christine Guillemot​​​‌ is co-supervising Soheib Takhtardeshir‌ together with Marten Sjostrom‌​‌ from MidSweden University in​​ the context of the​​​‌ Plenoptima Marie Curie project‌
• Christine Guillemot is co-supervising‌​‌ Samuel Willigham together with​​ Marten Sjostrom from MidSweden​​​‌ University in the context‌ of the Plenoptima Marie‌​‌ Curie project
• Christine Guillemot​​ is co-supervising Ipek Anil​​​‌ Atalay Appak together with‌ Humeyra Caglayan from Tampere‌​‌ University in the context​​ of the Plenoptima Marie​​​‌ Curie project
• Christine Guillemot‌ is co-supervising Leo-Paul Huar‌​‌ together with Pierre Hellier​​ in the context of​​​‌ a Cifre contract with‌ InterDigital.
• Nicolas Keriven and‌​‌ Aline Roumy are co-supervising​​ the PhD thesis of​​​‌ Antonin Joly in the‌ context of the PEPR‌​‌ SHARP
• Nicolas Keriven and​​ Aline Roumy are co-supervising​​​‌ the PhD thesis of‌ Adarsh Jamadandi in the‌​‌ context of the ERC​​ MALAGA
• Aline Roumy is​​​‌ co-supervising the PhD thesis‌ of Antoine Monier with‌​‌ Pierre Hellier in the​​ context of the joint​​​‌ Inria/InterDigital research project (defi‌ commun) Nisk.AI.

11.2.2 Juries‌​‌

• Christine Guillemot was member,​​​‌ as chair, of the​ PhD jury of Shubhendu​‌ JENA of the University​​ of Rennes, June 2025.​​​‌
• Christine Guillemot was member,​ as rapporteur, of the​‌ PhD jury of Aytaç​​ Özkan at the Technical​​​‌ University of Berlin, Dec.​ 2025.
• Thomas Maugey was​‌ member, as examiner, of​​ the PhD jury of​​​‌ Goluck KONUKO at the​ Paris-Saclay University, Jan. 2025.​‌
• Thomas Maugey was member,​​ as President, of the​​​‌ PhD jury of Sébastien​ DAM at the Rennes​‌ University, Oct. 2025.
• Thomas​​ Maugey was member, as​​​‌ rapporteur, of the PhD​ jury of Gabriele SPADARO​‌ at TELECOM Paris Institut​​ Polytechnique, Dec. 2025.
• Aline​​​‌ Roumy was member of​ the PhD committee of​‌ Corentin Presvôts, Paris-Saclay University,​​ Jan. 2025, as a​​​‌ chair.
• Aline Roumy was​ member of the PhD​‌ committee of Rodrigo Borba​​ Pinheiro, Paris-Saclay University, Jan.​​​‌ 2025, as a chair.​
• Aline Roumy was member​‌ of the PhD committee​​ of Jeremy Jaspar, Sorbonne​​​‌ Paris-Nord University, March. 2025,​ as a reviewer.
• Aline​‌ Roumy was member of​​ the PhD committee of​​​‌ Pierre-Alain.Afro, Grenoble Alpes University,​ April. 2025, as a​‌ reviewer.
• Aline Roumy was​​ member of the PhD​​​‌ committee of Maxime Ossonce,​ Paris-Saclay University, Dec. 2025,​‌ as an examiner.

11.2.3​​ Internal or external Inria​​​‌ responsibilities

• Aline Roumy is​ a member of the​‌ Gender Equality committee of​​ Inria-Rennes and Irisa, responsible​​​‌ for the working group​ on career interruptions and​‌ support.
• Aline Roumy is​​ a member of the​​​‌ mentorship program as a​ mentor.
• Thomas Maugey is​‌ a member of the​​ Formation Spécialisée de Site​​​‌, responsible of the​ security at work
• Thomas​‌ Maugey is a member​​ of the SEnS group,​​​‌ animating the reflexion on​ our research goals and​‌ impacts at the level​​ of the laboratory.

11.3.1 Specific official​ responsibilities in science outreach​‌ structures

• Thomas Maugey is​​ Scientific mediation officer in​​​‌ the scientific mediation team​ of Inria centre at​‌ Rennes Universiy.

11.3.2 Productions​​ (articles, videos, podcasts, serious​​​‌ games, ...)

• Thomas Maugey​ is the co-designer and​‌ co-supervisor of the project​​ Ma thèse une sacré​​​‌ histoire

11.3.3 Participation in​ Live events

• Thomas Maugey​‌ attended the "scientific mediation​​ days of Inria" at​​​‌ the Ministère de l'enseignement​ supérieur et de la​‌ recherche, and did a​​ presentation on the ma​​​‌ thèse une sacré histoire​ project.

International journals

• 5‌ articleN.Nicolas Charpenay‌​‌, M. L.Maël​​ Le Treust and A.​​​‌Aline Roumy. On‌ the Additivity of Optimal‌​‌ Rates for Independent Zero-Error​​ Source and Channel Problems​​​‌.IEEE Transactions on‌ Information Theory2025,‌​‌ 1-1HALDOIback​​ to textback to​​​‌ text
• 6 articleD.‌ R.Davi R Freitas‌​‌, I.Ioan Tabus​​ and C.Christine Guillemot​​​‌. Visibility-Based Geometry Pruning‌ of Neural Plenoptic Scene‌​‌ Representations.IEEE Transactions​​ on Multimedia2025,​​​‌ 1-17In press. HAL‌DOI
• 7 articleM.‌​‌Martin Gjorgjevski, N.​​Nicolas Keriven, S.​​​‌Simon Barthelme and Y.‌ D.Yohann De Castro‌​‌. Node Regression on​​ Latent Position Random Graphs​​​‌ via Local Averaging.‌Journal of Machine Learning‌​‌ Research2025. In​​ press. HALback to​​​‌ text
• 8 articleS.‌Soheib Takhtardeshir, R.‌​‌Roger Olsson, C.​​Christine Guillemot and M.​​​‌Mårten Sjöström. DUALF-D:‌ Disentangled dual-hyperprior approach for‌​‌ light field image compression​​.Signal Processing: Image​​​‌ Communication140January 2026‌, 117436HALDOI‌​‌back to text
• 9​​ articleS.Soheib Takhtardeshir​​​‌, R.Roger Olsson‌, C.Christine Guillemot‌​‌ and M.Mårten Sjöström​​. Efficient and Fast​​​‌ Light Field Compression via‌ VAE-Based Spatial and Angular‌​‌ Disentanglement.IEEE Access​​132025, 18594-18607​​​‌HALDOIback to‌ text

International peer-reviewed conferences‌​‌

• 10 inproceedingsS.Sara​​ Al Sayyed, A.​​​‌Aline Roumy and T.‌Thomas Maugey. Efficient‌​‌ Constraining of Transcoding in​​ DNA-Based Image Storage.​​​‌ICIP 2025 - IEEE‌ International Conference on Image‌​‌ ProcessingAnchorage (AK), USA,​​ United StatesIEEE2025​​​‌, 1-6HALback‌ to text
• 11 inproceedings‌​‌S.Sara Al Sayyed​​, A.Aline Roumy​​​‌, T.Thomas Maugey‌, N.Nicolas Lobato-Dauzier‌​‌ and A. J.Anthony​​ J Genot. Compact​​​‌ image representation for content-based‌ image retrieval in DNA‌​‌ data storage.Picture​​ Coding Symposium (PCS)PCS​​​‌ 2025 - Picture Coding‌ SymposiumAachen, Germany2025‌​‌, 1-5HALback​​ to text
• 12 inproceedings​​​‌A.Adarsh Jamadandi,‌ J.Jing Xu,‌​‌ A.Adam Dziedzic and​​ F.Franziska Boenisch.​​​‌ Memorization in Graph Neural‌ Networks.NeurIPS 2025‌​‌ - 39th Conference on​​ Neural Information Processing Systems​​​‌San Diego (CA), United‌ States2025HAL
• 13‌​‌ inproceedingsA.Antonin Joly​​, N.Nicolas Keriven​​​‌ and A.Aline Roumy‌. Taxonomy of reduction‌​‌ matrices for Graph Coarsening​​.Advances in Neural​​​‌ Information Processing Systems (NeurIPS)‌NeurIPS 2025 - 39th‌​‌ Annual Conference on Neural​​ Information Processing SystemsSan​​​‌ diego, CA, United States‌December 2025HALback‌​‌ to textback to​​ text
• 14 inproceedingsR.​​​‌Reda Kaafarani, J.‌ L.Julien Le Tanou‌​‌, M.Michael Ropert​​, T.Thomas Maugey​​​‌ and A.Aline Roumy‌. Joint multi-profile video‌​‌ coding for cost optimization​​ of standard ABR streaming​​​‌.ACM Mile-High Video‌ Conference (MHV 2025)MHV‌​‌ 2025 - 4th ACM​​​‌ Mile-High Video ConferenceDenver​ (CO), United StatesFebruary​‌ 2025, 33 -​​ 39HALDOI
• 15​​​‌ inproceedingsE.Esteban Pesnel​, J.Julien Le​‌ Tanou, M.Michael​​ Ropert, T.Thomas​​​‌ Maugey and A.Aline​ Roumy. SCALED :​‌ Surrogate-gradient for Codec-Aware Learning​​ of Downsampling in ABR​​​‌ Streaming.Picture Coding​ Symposium (PCS)PCS 2025​‌ - Picture Coding Symposium​​Aachen (Aix la Chapelle),​​​‌ GermanyDecember 2025HAL​back to text
• 16​‌ inproceedingsD.David Stéphane​​ Belemkoabga, T.Thomas​​​‌ Maugey and C.Christine​ Guillemot. GS-Morph: Dynamic​‌ Novel View Synthesis via​​ UDF-ARAP Gaussian Splat Morphing​​​‌.CVMP 2025 -​ ACM SIGGRAPH European Conference​‌ on Visual Media Production​​London, United KingdomACM​​​‌2025, 1-10HAL​DOIback to text​‌back to text
• 17​​ inproceedingsS.Soheib Takhtardeshir​​​‌, R.Roger Olsson​, C.Christine Guillemot​‌ and M.Mårten Sjöström​​. DUALF-C: Disentangled Light​​​‌ Field Compression with Entropy-Aware​ Bitstream Generation.VCIP​‌ 2025 - International Conference​​ on Visual Communications and​​​‌ Image ProcessingKlagenfurt, Austria​2025, 1-5HAL​‌back to text
• 18​​ inproceedingsP.Paul Wawerek-López​​​‌, N.Navid Mahmoudian​ Bidgoli, P.Pascal​‌ Frossard, A.Andre​​ Kaup and T.Thomas​​​‌ Maugey. OSLO-IC: On-the-Sphere​ Learned Omnidirectional Image Compression​‌ with Attention Modules and​​ Spatial Context.ICASSP​​​‌ 2025 - IEEE International​ Conference on Acoustics, Speech,​‌ and Signal ProcessingHyderabad,​​ India2025, 1-4​​​‌HALback to text​
• 19 inproceedingsS.Samuel​‌ Willingham, M.Mårten​​ Sjöström and C.Christine​​​‌ Guillemot. Maximum A​ Posteriori Training of Diffusion​‌ Models for Image Restoration​​.EUSIPCO 2025 –​​​‌ 33rd European Signal Processing​ ConferenceEUSIPCO 2025 -​‌ 33rd European Signal Processing​​ ConferenceIsola delle Femmine,​​​‌ Italy2025, 1-5​HALback to text​‌

National peer-reviewed Conferences

• 20​​ inproceedingsM.Martin Gjorgjevski​​​‌, N.Nicolas Keriven​, S.Simon Barthelme​‌ and Y. D.Yohann​​ De Castro. Graphical​​​‌ Kernel Ridge Regression in​ Latent Position Models.​‌GRETSI 2025 - 30e​​ Colloque Francophone sur le​​​‌ Traitement du Signal et​ des ImagesGRETSI 2025​‌ - 30e Colloque Francophone​​ sur le Traitement du​​​‌ Signal et des Images​Strasbourg, France2025,​‌ 1-4HAL
• 21 inproceedings​​A.Antonin Joly,​​​‌ N.Nicolas Keriven and​ A.Aline Roumy.​‌ Matrices de réduction et​​ de reconstruction pour la​​​‌ réduction de graphes.​GRETSI 2025 - 30ème​‌ Colloque Francophone de Traitement​​ du Signal et des​​​‌ ImagesStrasbourg, France2025​, 1-4HAL
• 22​‌ inproceedingsT.Thomas Maugey​​, D.Daniel Menard​​​‌, A.-C.Anne-Cécile Orgerie​ and R.Robin Richard​‌. Diffusion vidéo :​​ dépasser l'objectif d'efficacité et​​​‌ viser une sobriété choisie​.GRETSI - XXXe​‌ Colloque Francophone de Traitement​​ du Signal et des​​​‌ ImagesGRETSI 2025 -​ XXXe Colloque Francophone de​‌ Traitement du Signal et​​ des ImagesStrasbourg, France​​​‌2025, 1-4HAL​back to text

Doctoral​‌ dissertations and habilitation theses​​

• 23 thesisT.Tom​​​‌ Bordin. Image semantic​ compression at extremely low​‌ bitrates.Université de​​ Rennes 1December 2025​​HAL
• 24 thesisD.​​​‌ R.Davi Rabbouni de‌ Carvalho Freitas. Sensing‌​‌ and reconstruction of plenoptic​​ point clouds.Université​​​‌ de Tampere (Finlande)May‌ 2025HAL
• 25 thesis‌​‌R.Rémi Piau.​​ Learning from entropy-encoded data​​​‌.Université de Rennes‌June 2025HALback‌​‌ to text

Reports &​​ preprints

• 26 miscH.​​​‌Hugo Jaquard and N.‌Nicolas Keriven. Statistical‌​‌ Consistency of Discrete-to-Continuous Limits​​ of Determinantal Point Processes​​​‌.2026HAL
• 27‌ misc N.Nicolas Keriven‌​‌. Backward Oversmoothing: why​​ is it hard to​​​‌ train deep Graph Neural‌ Networks? 2025 HAL back‌​‌ to text
• 28 misc​​C.Caroline Mazini Rodrigues​​​‌, N.Nicolas Keriven‌ and T.Thomas Maugey‌​‌. Compressing image encoders​​ via latent distillation.​​​‌2026HALback to‌ text