Answers are grounded in this paper and cite their sources.
Objects that reappear after occlusion can receive new identities when data association fails, fragmenting trajectories. We present ReSORT, an identity recovery framework that evaluates competing links using motion history, numerical association scores, and selective language-model evidence. Camera-compensated trajectory states are propagated through five motion hypotheses to construct a persistent candidate graph. A numerical ranker scores candidate links, and a frozen Qwen3.5-4B model reviews selected ambiguous cases using image crops and scene context. Candidate-order consistency checks filter the model responses before their relative evidence enters a global one-to-one assignment. Inference does not access ground truth; the numerical ranker is trained with supervision, and the evaluated reassociation uses subsequent observations offline. On seven MOT17 development videos, the complete pipeline produces 243 identity switches, compared with 427 for native ByteTrack and 299 for the historical automatic ReSORT reference, with IDF1 of 85.96% and HOTA of 78.06%. The historical comparison includes changes beyond the final language-model component, whose contribution is assessed separately through matched ablations. Experiments across seven tracker families on MOT17 and MOT20 show tracker-dependent benefits and sensitivity to evidence weighting. The results support selective language-model verification within a constrained association system while identifying limitations in response consistency, computational cost, and development-set evaluation.
No takes yet. Share an insight, caveat, or question.
roh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: