Unsupervised Domain Adaptation for Pose Estimation and Tracking

• Auteur / Autrice : Melek Elloumi
• Direction : Faten Chaieb-Chakchouk
• Type : Projet de thèse
• Discipline(s) : Informatique
• Date : Inscription en doctorat le 01/09/2023
• Etablissement(s) : Université Paris-Panthéon-Assas
• Ecole(s) doctorale(s) : École doctorale des sciences économiques et gestion, sciences de l'information et de la communication (Paris)

Résumé

This thesis will focus on developing novel Unsupervised Domain Adaptation (UDA) techniques for generic pose estimation to bridge the performance gap due to domain shifts. This implies novel fundamental contributions on UDA for regression problems. Applications of interest include but are not limited to animal pose estimation and landmark detection and tracking for sports applications. The sub-objectives of the thesis can be summarized to: • Understand, quantify, and analyse the domain shift problem for existing pose estimation techniques; • Investigate and propose novel fundamental UDA approaches for regression problems; • Apply, test, and validate the proposed UDA techniques for the problem of horse pose estimation on real-world scenarios; • Apply, test, and validate the proposed UDA techniques for the problem of landmark detection and tracking in sports (e.g., Badminton, Tennis, Tennis Table, etc.).