Thomas Guionnet
Research Subjects
Thomas Guionnet has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Mickael Raulet
ATEME
9 shared publications - Thomas Burnichon
ATEME, Rennes, France
4 shared publications - Marc Rivière
ATEME
3 shared publications - Marwa Tarchouli
ATEME
2 shared publications - Nikolay Tverdokhleb
ATEME, Rennes, France
2 shared publications - Abdelmajid Moussaoui1 shared publication
Latest Publications
- Open Access
- 18 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 40-46
NEURAL VIDEO COMPRESSION OVERVIEW, PERFORMANCE AND CHALLENGES
- 18 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 40-46
Open AccessNEURAL VIDEO COMPRESSION OVERVIEW, PERFORMANCE AND CHALLENGES
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 14-20
End-to-end VVC encoding for energy-aware HDR rendering
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 14-20
Open AccessEnd-to-end VVC encoding for energy-aware HDR rendering
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 85-86
AV1-Compatible Film Grain Modeling for HEVC Decoders
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 85-86
Open AccessAV1-Compatible Film Grain Modeling for HEVC Decoders
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 98
AI-based monocular depth map estimation applied to a video encoding pipeline
Depth information is used in many applications, including computational photography, especially in smartphones, scene analysis and 3D reconstruction. Obviously, depth information is useful for 3D related applications. It may seem less related to 2D applications such as traditional 2D video compression. However, it is demonstrated in this paper that depth information is a practical and useful addition to a video compression pipeline. In a first part, the problem of depth map computation is considered. For a long time, depth estimation required specific image features such as stereo pairs, specific camera motion or particular structures in the captured scene. More recently though, artificial intelligence (AI) unlocked fully generic depth estimation in monocular images. Thanks to careful algorithmic and implementation optimization, a real-time depth map computation method is proposed. In a second part, the usefulness of depth map estimation is considered through the example of frame interpolation, illustrated by experimental results. Finally it is discussed how depth maps can be an elementary analysis serving several steps in an encoding pipeline.
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 98
Open AccessAI-based monocular depth map estimation applied to a video encoding pipeline
Depth information is used in many applications, including computational photography, especially in smartphones, scene analysis and 3D reconstruction. Obviously, depth information is useful for 3D related applications. It may seem less related to 2D applications such as traditional 2D video compression. However, it is demonstrated in this paper that depth information is a practical and useful addition to a video compression pipeline. In a first part, the problem of depth map computation is considered. For a long time, depth estimation required specific image features such as stereo pairs, specific camera motion or particular structures in the captured scene. More recently though, artificial intelligence (AI) unlocked fully generic depth estimation in monocular images. Thanks to careful algorithmic and implementation optimization, a real-time depth map computation method is proposed. In a second part, the usefulness of depth map estimation is considered through the example of frame interpolation, illustrated by experimental results. Finally it is discussed how depth maps can be an elementary analysis serving several steps in an encoding pipeline.
- Open Access
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 87
MV-HEVC
Multiview High Efficiency Video Coding (MV-HEVC) is an HEVC extension focused on efficiently coding spatially related images, such as a left eye and right eye views of 3D stereoscopic content. MV-HEVC was released in the second version of HEVC back in October 2014 [2]. MV-HEVC was created alongside another extension named SHVC, which is already used in ATSC 3.0 use cases since 2018 [1]. The two extensions share many fundamentals, codified in an annex for common specifications for multi-layer extensions. Their base principle is scalability: A traditional HEVC decoder can rely on the base layer only, hence providing backward compatibility, but a more elaborate decoder can use the second layer as well to improve the rendition. The MV-HEVC format recently gained widespread attention with Apple announcing support for 3D movies on the Apple Vision Pro at WWDC23 [9][8]. Apple provided public guidelines on stereo video encoding [3], packaging [5], and streaming [4] as well as on providing parallax information to improve caption placement [6]. Apple also organized Apple Vision Pro developer labs [7]. This presentation will describe the MV-HEVC standard and how it works in theory. An implementation of MV-HEVC [10] will also be studied to assess performance in practice in industrial settings. Deployment aspects such as packaging and captioning will be discussed. Finally, 3D-specific optimization strategies will be analyzed, including software architecture and rate control.
- 11 February 2024
- Published by Association for Computing Machinery (ACM)
- p. 87
Open AccessMV-HEVC
Multiview High Efficiency Video Coding (MV-HEVC) is an HEVC extension focused on efficiently coding spatially related images, such as a left eye and right eye views of 3D stereoscopic content. MV-HEVC was released in the second version of HEVC back in October 2014 [2]. MV-HEVC was created alongside another extension named SHVC, which is already used in ATSC 3.0 use cases since 2018 [1]. The two extensions share many fundamentals, codified in an annex for common specifications for multi-layer extensions. Their base principle is scalability: A traditional HEVC decoder can rely on the base layer only, hence providing backward compatibility, but a more elaborate decoder can use the second layer as well to improve the rendition. The MV-HEVC format recently gained widespread attention with Apple announcing support for 3D movies on the Apple Vision Pro at WWDC23 [9][8]. Apple provided public guidelines on stereo video encoding [3], packaging [5], and streaming [4] as well as on providing parallax information to improve caption placement [6]. Apple also organized Apple Vision Pro developer labs [7]. This presentation will describe the MV-HEVC standard and how it works in theory. An implementation of MV-HEVC [10] will also be studied to assess performance in practice in industrial settings. Deployment aspects such as packaging and captioning will be discussed. Finally, 3D-specific optimization strategies will be analyzed, including software architecture and rate control.
- Open Access
CNN-based Prediction of Partition Path for VVC Fast Inter Partitioning Using Motion Fields
The Versatile Video Coding (VVC) standard has been recently finalized by the Joint Video Exploration Team (JVET). Compared to the High Efficiency Video Coding (HEVC) standard, VVC offers about 50% compression efficiency gain, in terms of Bjontegaard Delta-Rate (BD-rate), at the cost of a 10-fold increase in encoding complexity. In this paper, we propose a method based on Convolutional Neural Network (CNN) to speed up the inter partitioning process in VVC. Firstly, a novel representation for the quadtree with nested multi-type tree (QTMT) partition is introduced, derived from the partition path. Secondly, we develop a U-Net-based CNN taking a multi-scale motion vector field as input at the Coding Tree Unit (CTU) level. The purpose of CNN inference is to predict the optimal partition path during the Rate-Distortion Optimization (RDO) process. To achieve this, we divide CTU into grids and predict the Quaternary Tree (QT) depth and Multi-type Tree (MT) split decisions for each cell of the grid. Thirdly, an efficient partition pruning algorithm is introduced to employ the CNN predictions at each partitioning level to skip RDO evaluations of unnecessary partition paths. Finally, an adaptive threshold selection scheme is designed, making the trade-off between complexity and efficiency scalable. Experiments show that the proposed method can achieve acceleration ranging from 16.5% to 60.2% under the RandomAccess Group Of Picture 32 (RAGOP32) configuration with a reasonable efficiency drop ranging from 0.44% to 4.59% in terms of BD-rate, which surpasses other state-of-the-art solutions. Additionally, our method stands out as one of the lightest approaches in the field, which ensures its applicability to other encoders.
Open AccessCNN-based Prediction of Partition Path for VVC Fast Inter Partitioning Using Motion Fields
The Versatile Video Coding (VVC) standard has been recently finalized by the Joint Video Exploration Team (JVET). Compared to the High Efficiency Video Coding (HEVC) standard, VVC offers about 50% compression efficiency gain, in terms of Bjontegaard Delta-Rate (BD-rate), at the cost of a 10-fold increase in encoding complexity. In this paper, we propose a method based on Convolutional Neural Network (CNN) to speed up the inter partitioning process in VVC. Firstly, a novel representation for the quadtree with nested multi-type tree (QTMT) partition is introduced, derived from the partition path. Secondly, we develop a U-Net-based CNN taking a multi-scale motion vector field as input at the Coding Tree Unit (CTU) level. The purpose of CNN inference is to predict the optimal partition path during the Rate-Distortion Optimization (RDO) process. To achieve this, we divide CTU into grids and predict the Quaternary Tree (QT) depth and Multi-type Tree (MT) split decisions for each cell of the grid. Thirdly, an efficient partition pruning algorithm is introduced to employ the CNN predictions at each partitioning level to skip RDO evaluations of unnecessary partition paths. Finally, an adaptive threshold selection scheme is designed, making the trade-off between complexity and efficiency scalable. Experiments show that the proposed method can achieve acceleration ranging from 16.5% to 60.2% under the RandomAccess Group Of Picture 32 (RAGOP32) configuration with a reasonable efficiency drop ranging from 0.44% to 4.59% in terms of BD-rate, which surpasses other state-of-the-art solutions. Additionally, our method stands out as one of the lightest approaches in the field, which ensures its applicability to other encoders.
- Open Access
- 7 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 14-20
Perceptual quality evaluation of variable resolution VVC video coding using RPR
In this paper we analyze the perceptual performance of encoding in reduced resolution at challenging bitrates on 4k content by means of reference picture resampling tool in VVC in comparison with encoding in source resolution. The subjective testing shows that adaptively encoding in reduced resolution based on GOP-based reference picture resampling encoder control present in the VVC reference software is always better or equal to the anchor given a 95% confidence interval. The objective metrics based on PSNR and MS-SSIM also both indicate a gain for these sequences.
- 7 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 14-20
Open AccessPerceptual quality evaluation of variable resolution VVC video coding using RPR
In this paper we analyze the perceptual performance of encoding in reduced resolution at challenging bitrates on 4k content by means of reference picture resampling tool in VVC in comparison with encoding in source resolution. The subjective testing shows that adaptively encoding in reduced resolution based on GOP-based reference picture resampling encoder control present in the VVC reference software is always better or equal to the anchor given a 95% confidence interval. The objective metrics based on PSNR and MS-SSIM also both indicate a gain for these sequences.
- Open Access
- 7 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 78-84
AI-BASED LIGHT PARALLEL VIDEO ENCODER
This paper presents a software framework for end-to-end learned video coding. The main goal is to provide the research community with a platform adapted to practical applications. To accomplish this purpose, several features have been integrated, including basic utilities such as color conversion and advanced functions such as overlapping patch-coding to enable flexible memory consumption and adaptation to various hardware parallelization capabilities. The MS-SSIM function has been altered to fit this approach, and the rationale behind this modification is explained. This framework is meant to serve as a basis for validation and comparison of state-of-the-art end-to-end learned video codecs. Therefore, the codec part is a module that can be replaced while keeping the framework structure, and performance metrics are included. Experimental results are provided to illustrate the effectiveness of the framework.
- 7 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 78-84
Open AccessAI-BASED LIGHT PARALLEL VIDEO ENCODER
This paper presents a software framework for end-to-end learned video coding. The main goal is to provide the research community with a platform adapted to practical applications. To accomplish this purpose, several features have been integrated, including basic utilities such as color conversion and advanced functions such as overlapping patch-coding to enable flexible memory consumption and adaptation to various hardware parallelization capabilities. The MS-SSIM function has been altered to fit this approach, and the rationale behind this modification is explained. This framework is meant to serve as a basis for validation and comparison of state-of-the-art end-to-end learned video codecs. Therefore, the codec part is a module that can be replaced while keeping the framework structure, and performance metrics are included. Experimental results are provided to illustrate the effectiveness of the framework.
- 1 January 2023
- Published by Society of Motion Picture and Television Engineers (SMPTE)
- p. 16-22
The future of video compression - Moving beyond hybrid codecs with machine learning
The consumption of video content on the internet is increasing at a constant pace, along with an increase of video quality. As an answer to the ever-growing demand for high quality video, compression technology improves steadily. About every decade, a new major video compression standard is issued, providing a decrease of bitrate by a factor two. Interestingly, the technology does not change radically between codecs generations. Instead, the same principles are re-used and pushed further. There has been several attempts to depart from this model, but none achieved to be competitive. Recently, the research community has started focusing on deep learning-based strategies. Could it be the new contender to the classical approach? This paper analyzes the benefits and limitations of deep learning-based video compression methods, and investigates practical aspects such as rate control, delay, memory consumption and power consumption. Overlapping patch-based end-to-end video compression strategy is proposed to overcome memory limitations.
- 1 January 2023
- Published by Society of Motion Picture and Television Engineers (SMPTE)
- p. 16-22
The future of video compression - Moving beyond hybrid codecs with machine learning
The consumption of video content on the internet is increasing at a constant pace, along with an increase of video quality. As an answer to the ever-growing demand for high quality video, compression technology improves steadily. About every decade, a new major video compression standard is issued, providing a decrease of bitrate by a factor two. Interestingly, the technology does not change radically between codecs generations. Instead, the same principles are re-used and pushed further. There has been several attempts to depart from this model, but none achieved to be competitive. Recently, the research community has started focusing on deep learning-based strategies. Could it be the new contender to the classical approach? This paper analyzes the benefits and limitations of deep learning-based video compression methods, and investigates practical aspects such as rate control, delay, memory consumption and power consumption. Overlapping patch-based end-to-end video compression strategy is proposed to overcome memory limitations.
- 1 May 2022
- Published by Society of Motion Picture and Television Engineers (SMPTE) in SMPTE Motion Imaging Journal
- Vol. 131 (4) , 45-49
Dynamic Seamless Resource Allocation for Live Video Compression on a Kubernetes Cluster
A solution is proposed on top of Kubernetes to dynamically allocate service resources without service interruption. It is the basis for optimizing a live video compression service. It is demonstrated that dynamic resource allocation can benefit a video compression application, either by reducing the resource consumption, hence costs, or by enhancing delivered video quality. By combining the proposed solution with an elastic encoder and machine learning (ML) for content complexity estimation, a content- and application-aware dynamic resource orchestrator for realtime video compression is designed. Preliminary experimental results using ATEME Titan Live Micro-Services 1 encoders demonstrate substantial bitrate reductions on even the most demanding channel.
- 1 May 2022
- Published by Society of Motion Picture and Television Engineers (SMPTE) in SMPTE Motion Imaging Journal
- Vol. 131 (4) , 45-49
Dynamic Seamless Resource Allocation for Live Video Compression on a Kubernetes Cluster
A solution is proposed on top of Kubernetes to dynamically allocate service resources without service interruption. It is the basis for optimizing a live video compression service. It is demonstrated that dynamic resource allocation can benefit a video compression application, either by reducing the resource consumption, hence costs, or by enhancing delivered video quality. By combining the proposed solution with an elastic encoder and machine learning (ML) for content complexity estimation, a content- and application-aware dynamic resource orchestrator for realtime video compression is designed. Preliminary experimental results using ATEME Titan Live Micro-Services 1 encoders demonstrate substantial bitrate reductions on even the most demanding channel.