VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking with Highly Similar Objects

Multi-object tracking (MOT) often struggles when objects are small, densely distributed, and visually indistinguishable, such as in bee swarms. Traditional re-identification (re-ID) models are trained to match individual object instances, but MOT inference requires global association between multiple historical trajectories and current-frame detections. This paper investigates whether aligning the training process with this video-level association requirement can reduce identity switches and improve tracking accuracy.

The authors propose Video-Level Association re-ID (VLA-ReID), which reformulates the re-ID task to match the global association process used during inference. Instead of training on isolated pairs, the model is trained on video-level episodes that present historical trajectory features as queries against all current-frame detections. To handle high appearance similarity, the authors introduce two modules: Frame-Common Appearance Estimation (FCAE) and Common-Appearance Suppression (CAS). FCAE identifies the dominant appearance features shared by all detections in a frame, and CAS suppresses these common components in both trajectory and detection features. This forces the model to focus on the residual, discriminative differences between objects without requiring additional annotations.

Experiments on the BEE24 dataset demonstrate that VLA-ReID significantly improves tracking performance compared to state-of-the-art trackers. The method achieved notable gains in key metrics, including HOTA (+1.1), AssR (+2.6), and IDF1 (+0.8). Most importantly, the approach reduced identity switches by 28%, confirming that video-level association modeling and common-appearance suppression effectively mitigate the ambiguity inherent in tracking visually similar objects.