Public Thesis Defense of Vladimir Somers - ICTEAM

Person Re-Identification and its Application to Multi-Object Tracking by Vladimir Somers

Le lundi 5 mai 2025 à 16h15 – Salle Shannon (bâtiment Maxwell), place du Levant à 1348 Louvain-la-Neuve

Multi-Object Tracking (MOT) is a fundamental computer vision task that involves detecting objects of interest in video frames and associating detections of the same object across time to form trajectories. For association, appearance cues, extracted through person re-identification (ReID) models, play a crucial role by capturing distinctive visual features of the tracked targets. However, despite its importance for tracking, ReID has primarily been studied as an image retrieval problem, with state-of-the-art methods overlooking tracking-specific challenges. This thesis focuses on advancing person re-identification methods, with a particular emphasis on making them more robust and better suited for tracking applications. A key challenge in ReID and MOT is handling occlusions, where targets become partially hidden by objects or other people, leading to degraded re-identification accuracy and potential identity switches in tracking. Additionally, tracking methods often fail to effectively combine ReID with motion cues and scene context, relying instead on naive association strategies. To address these challenges, this thesis makes three key contributions: BPBreID, a part-based method for robust occluded re-identification; KPR, a keypoint-prompted ReID model designed to address multi-person occlusions scenarios; and CAMELTrack, an online tracking-by-detection method that replaces traditional heuristic for detection association with a context-aware learnable module. At the time of writing, KPR and CAMELTrack achieve state-of-the-art performance on widely-used benchmarks for occluded person re-identification and multi-object tracking. Beyond these core contributions, I also contributed to sports analytics by developing a specialized re-identification model for sports players and introducing novel datasets for player tracking, re-identification, and jersey number recognition. The thesis concludes with a critical analysis of the current state of tracking and re-identification technologies, offering my opiniated view about the future of these two rapidly evolving fields.

Jury members :
Prof. Christophe De Vleeschouwer (UCLouvain), supervisor
Prof. Alexandre Alahi (EPFL, Suisse), supervisor
Prof. Laurent Francis (UCLouvain), chairperson
Prof. Laurent Jacques (UCLouvain), secretary
Prof. Andrea Cavallaro (EPFL, Suisse)
Prof. Laura Leal-Taixé (NVIDIA)

Pay attention: the public defense of Vladimir Somers will also take place in the form of a videoconference.