Titre de la thèse
High-Order Statistics for Image Representations using Metric Learning
Date et lieu de soutenance
Mardi 8 septembre 2020, 14h.
ENSEA Cergy, salle à préciser.
An important challenge in artificial intelligence is the learning of useful data representations, or features, on which recognition algorithms can be used in order to solve a specific task (object or scene recognition, video annotation, etc.). By learning rich representations, that is, which encode the relevant information with respect to the given task, performances of the majority of recognition algorithms can be highly improved. The major difficulty in learning such representations rests on the high dimensionality of the input data, and the lack of prior knowledge about the relevant patterns that should be encoded.
In this thesis, we particularly focus on image representation for the task of content-based image retrieval using the deep metric learning framework. These representations must be low-dimensional in order to reduce the memory storage, the computation cost, and to speed-up the search. Also, they should encode useful information in order to be able to retrieve the most relevant images from the database.
We make the following contributions to alleviate the aforementioned issues: First, we present three image representations that leverage dictionary learning in order to produce richer representation along with covariance matrices or attention mechanisms. Second, we propose two improvements during the training: a method that generates hard examples to resume the training, and a regularization that improves the robustness of the local features for a variety of representations. We empirically show that the proposed strategies significantly improve baseline methods and provide stronger results than most of the state-of-the-art methods.
Composition du jury
- Mme LARLUS Diane, Senior Research Scientist, Navers Labs Europe, Examinatrice
- Mme VINCENT Nicole, Professeure des Universités, Université de Paris, Examinatrice
- M. JEGOU Hervé, Senior Research Scientist, Facebook AI, Examinateur
- M. CHUM Ondrej, Associate Professor, Czech Technical University in Prague, Rapporteur
- M.CORD Matthieu, Professeur des Universités, Sorbonne Université, Rapporteur
- M. KLEIN Edouard, Capitaine, Pôle Judiciaire de la Gendarmerie Nationale, Co-encadrant de thèse
- M. PICARD David, Senior Research Scientist, Ecole des Ponts ParisTech, Co-encadrant de thèse
- M. HISTACE Aymeric, Professeur des Universités, ENSEA, Directeur de thèse