Title: Seg2Track++: Enhancing Multi-Object Tracking and Segmentation via Probabilistic Track Validation and Data Association

Abstract:

Reliable operation of autonomous systems in dynamic settings depends heavily on robust Multi-Object Tracking and Segmentation (MOTS), which guarantees stable object identities and accurate pixel-level mask delineation. While foundation models like SAM2 exhibit impressive zero-shot generalization capabilities for segmentation tasks, their direct deployment in MOTS is hindered by inconsistent track association and the propagation of false positives. To address these challenges, we present Seg2Track++, a novel framework that combines instance segmentation with SAM2 and a new track management module to achieve zero-shot MOTS with superior temporal consistency. This approach utilizes Mask Centroid Distance (MCD) and Confidence-Aware Cost Modulation (CCM) for track association, while Probabilistic Track Validation (PTV) leverages a Bernoulli filter to verify track existence and eliminate ghost tracks. Evaluations on the KITTI MOTS dataset reveal that Seg2Track++ significantly improves identity preservation and minimizes false-positive propagation, delivering robust track management without the need for fine-tuning.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC