Seg2Track++: Probabilistic Track Validation and Data Association for Multi-Object Tracking and Segmentation

Researchers have introduced two new methods for multi-object tracking and segmentation, enhancing the capabilities of autonomous systems in dynamic environments.

The first method, Seg2Track++, is a framework that integrates instance segmentation with SAM2 and a novel track management module to perform zero-shot Multi-Object Tracking and Segmentation (MOTS) with enhanced temporal consistency^[1]. It uses Mask Centroid Distance (MCD) and Confidence-Aware Cost Modulation (CCM) for track association, and Probabilistic Track Validation (PTV) with a Bernoulli filter to validate track existence and suppress ghost tracks. Experimental results on KITTI MOTS demonstrate improved identity preservation and reduced false-positive propagation. Meanwhile, a second method, Track-Detection Link Prediction (TDLP), has been proposed for multi-object tracking, which learns association directly from data without handcrafted rules^[2]. TDLP is a tracking-by-detection method that performs per-frame association via link prediction between tracks and detections, and is designed primarily for geometric features such as bounding boxes. TDLP consistently surpasses state-of-the-art performance across both tracking-by-detection and end-to-end methods, with link prediction proving more effective than metric learning-based association, particularly when handling heterogeneous features.