CHEN Yifu
Ban lãnh đạo nghiên cứu : Matthieu CORD
Deep learning for visual semantic segmentation
With the proliferation of cameras and communication tools, the amount of visual data available is constantly increasing. With this data, many fascinating applications can be developed today, such as automated driving systems or computer-assisted medical diagnosis. It is therefore important to develop scientific and technological tools that enable high-performance automatic analysis of visual data. In this thesis, we are interested in Visual Semantic Segmentation, one of the high-level task that paves the way towards complete scene understanding. Specifically, it requires a semantic understanding at the pixel level. With the success of deep learning in recent years, semantic segmentation problems are being tackled using deep architectures. Typically, these approaches consist of three components: a deep network, a loss function, and an optimization process on an annotated dataset. In the first part, we focus on the construction of a more appropriate loss function for semantic segmentation. More precisely, we define a novel loss function by employing a semantic edge detection network. This loss imposes pixel-level predictions to be consistent with the ground truth semantic edge information, and thus leads to better-shaped segmentation results. In the second part, we address another important issue, namely, alleviating the need for training segmentation models with large amounts of fully annotated data. We propose a novel attribution method that identifies the most significant regions in an image considered by classification networks. We then integrate our attribution method into a weakly supervised segmentation framework. The semantic segmentation models can thus be trained with only image-level labeled data, which can be easily collected in large quantities. All models proposed in this thesis are thoroughly experimentally evaluated on multiple datasets and the results are competitive with the literature.
Bảo vệ luận án : 09/09/2020
Hội đồng giám khảo :
Mme. Catherine Achard (Sorbonne Université - ISIR) Examinatrice
M. Patrick Lambert (Université Savoie Mont Blanc - LISTIC) Rapporteur
M. Sébastien Lefèvre (Université Bretagne Sud - IRISA) Rapporteur
Mme. Camille Couprie (Facebook AI Research) Examinatrice
M. Frédéric Precioso (Université Côte d'Azur - I3S) Examinateur
M. Arnaud Dapogny (Datakalab) Examinateur
M. Matthieu Cord (Sorbonne Université - LIP6) Directeur de thèse
Bài báo khoa học 2019-2021
-
2021
- A. Douillard, Y. Chen, A. Dapogny, M. Cord : “Tackling Catastrophic Forgetting and Background Shift in Continual Semantic Segmentation”, (2021)
- A. Douillard, Y. Chen, A. Dapogny, M. Cord : “PLOP: Learning without Forgetting for Continual Semantic Segmentation”, CVPR, Nashville, United States (2021)
-
2020
- Y. Chen : “Apprentissage profond pour la segmentation sémantique d’images”, luận án, bảo vệ luận án 09/09/2020, ban lãnh đạo nghiên cứu Cord, Matthieu (2020)
- Y. Chen, A. Dapogny, M. Cord : “SEMEDA: Enhancing Segmentation Precision with Semantic Edge Aware Loss”, Pattern Recognition, vol. 108, pp. 107557, (Elsevier) (2020)
-
2019
- A. Saporta, Y. Chen, M. Blot, M. Cord : “REVE: Regularizing Deep Learning with Variational Entropy Bound”, 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, Province of China, pp. 1610-1614, (IEEE) (2019)
- Y. Chen, A. Saporta, A. Dapogny, M. Cord : “Delving Deep into Interpreting Neural Nets with Piece-Wise Affine Representation”, 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, Province of China, pp. 609-613, (IEEE) (2019)