Publication

Seminars

Three new exciting seminars will be given on June 7th, 2019. The program is the following :

Place : Telecom-ParisTech, 46 rue Barrault, Paris. Room B567.

9h45h Welcome coffee
10h00-11h15 “Deep learning for Super Resolution and Tracking”, by Gianni Franchi, Univ. Paris Sud
11h15-12h30 “Deformation models for image and video generation”, by Stéphane Lathuilière, Univ. Trento (Italy)
12h30-14h Lunch break
14h-15h15 “The Next Big Thing: From Systems to Deep Systems”, by Francesco Banterle, CNR Pisa (Italy)
15h15-16h Coffee, free discussion

Abstracts :

“Deep learning for Super Resolution and Tracking”
Cette présentation traitera de deux projets :
Le projet 1 : vise à mélanger des techniques de deep learning et de géostatistique pour faire de la super résolution d’images. L’objectif est d’accéder aux bons résultats du deep learning ainsi que l’incertitude de l’estimateur grâce la géostatistique.
Le projet 2 : vise à suivre des personnes dans une vidéo de foule extrêmement dense. N’ayant pas de base donnée annotée sur ce projet, nous proposerons une technique ou le réseau de neurones apprend tout seul. (En anglais : Self supervised learning).

“Deformation models for image and video generation”
Generating realistic images and videos has countless applications in different areas, ranging from photography technologies to e-commerce business.
Recently, deep generative approaches have emerged as effective techniques for generation tasks. In this talk, we will first present the problem of pose-guided person image generation. Specifically, given an image of a person and a target pose, a new image of that person in the target pose is synthesized. We will show that important body-pose changes affect generation quality and that specific feature map deformations lead to better images.
Then, we will present our recent framework for video generation. More precisely, our approach generates videos where an object in a source image is animated according to the motion of a driving video. In this task, we employ a motion representation based on keypoints that are learned in a self-supervised fashion. Therefore, our approach can animate any arbitrary object without using annotation or prior information about the specific object to animate.

“The Next Big Thing: From Systems to Deep Systems”
The main communities in Computer Science are all shifting from traditional algorithms towards deep-based algorithms where deep learning is extensivelyused to solve everyday problems. Although this is very attractive in terms of quality and speed,the days of end-to-end encoding are numbered because more than a network is needed to achieve a full task. This talk will show a traditional system for 3D reconstruction, how to make it deep, and the making of a from scratch deep system in which deep learning was in the loop from start to finish,

Poste

Open position: Associate/Assistant Professor in Video Analysis and Learning

Telecom ParisTech, a CS/EE school of Institut Polytechnique de Paris, is hiring an Associate Professor in Video Analysis and Learning. The position will be located in the Multimedia Team, within the Image, Data, Signal Department (IDS), and the LTCI laboratory.

The Multimedia team has a long activity in the domain of video and image coding and transmission. More recently, video analysis and learning activity have become more and more relevant for the team who runs now a regional study group about Machine and Deep Learning applications to Image and Video compression. The team has the target to expand its activity in this area, and several new and exciting research projects have just been launched, such as research programs in Deep Learning assisted video compression and Learning-based photographic quality evaluation. In this context, and to support the increasing activity of the team, a permanent position in video analysis and learning has been opened.

Applicants are expected to provide an outstanding academic research record and will be encouraged to advise PhD theses, supervise engineers and post-docs, while being actively involved in funded projects and in the activities of the Multimedia team. The teaching activities will take place in the engineer and master tracks at Telecom ParisTech and can be given in English.

Find here more information.

Apply via e-mail.

Publication

ICASSP papers

Three articles have been accepted into IEEE ICASSP :
1) S. Zheng, M. Cagnazzo, M. Kieffer. “CHANNEL IMPULSIVE NOISE MITIGATION FOR LINEAR VIDEO CODING SCHEMES”
2) L. Wang, A. Fiandrotti, A. Purica, G. Valenzise, M. Cagnazzo. “ENHANCING HEVC SPATIAL PREDICTION BY CONTEXT-BASED LEARNING”
3) P. Nikitin, M. Cagnazzo, J. Jung. “COMPRESSION IMPROVEMENT VIA REFERENCE ORGANIZATION FOR 2D-MULTIVIEW CONTENT”.
Congrats to Shuo, Li and Pavel.

Soutenance

Shuo Zheng’s Phd defense

Shuo Zheng’s PhD defense will take place at 5th February, 10 am, Amphi Opale at TélécomParisTech (46 rue Barrault, 75013 Paris).

Committee:

  • Mr François-Xavier Coudoux, Université Polytechnique Hauts-de-France, Referee
  • Mrs Aline Roumy, INRIA Rennes, Referee
  • Mr Jean-Marie Gorce, INSA Lyon, Examiner
  • Mr Marc Leny, Ektacom, Examiner
  • Mrs Michèle Wigger, TélécomParitech, Examiner, Jury’s Chair
  • Mr Marco Cagnazzo, TélécomParisTech, Advisor
  • Mr Michel Kieffer, Université de Paris-sud, Advisor

Title: Accounting for Channel Constraints in Joint Source-Channel Video Coding Schemes

Abstract: SoftCast based Linear Video Coding (LVC) schemes have been emerged in the last decade as a quasi analog joint-source-channel alternative to classical video coding schemes. Theoretical analyses have shown that analog coding is better than digital coding in a multicast scenario when the channel signal-to-noise ratios (C-SNR) dier among receivers. LVC schemes provide in such context a decoded video quality at dierent receivers proportional to their C-SNR. This thesis considers rst the channel precoding and decoding matrix design problem for LVC schemes under a per-subchannel power constraint. Such constraint is found, e.g., on Power Line Telecommunication (PLT) channels and is similar to per-antenna power constraints in multi-antenna transmission system. An optimal design approach is proposed, involving a multi-level water lling algorithm and the solution of a structured Hermitian Inverse Eigenvalue problem. Three lower-complexity alternative suboptimal algorithms are also proposed. Extensive experiments show that the suboptimal algorithms perform closely to the optimal one and can reduce signicantly the complexity. The precoding matrix design in multicast situations also has been considered. A second main contribution consists in an impulse noise mitigation approach for LVC schemes. Impulse noise identication and correction can be formulated as a sparse vector recovery problem. A Fast Bayesian Matching Pursuit (FBMP) algorithm is adapted to LVC schemes. Subchannels provisioning for impulse noise mitigation is necessary, leading to a nominal video quality decrease in absence of impulse noise. A phenomenological model (PM) is proposed to describe the impulse noise correction residual. Using the PM model, an algorithm to evaluate the optimal number of subchannels to provision is proposed. Simulation results show that the proposed algorithms signicantly improve the video quality when transmitted over channels prone to impulse noise.

Professional blog