ESC-SORT: enhanced similarity costs lead to an improved multi-object tracker

Le résumé fourni par la source

Comprehensive object detection and correct data association are the most critical tasks in the two-stage paradigm of multi-object tracking (MOT). For data association, most existing trackers resort to appearance features and complicated motion models, causing the methods to either suffer from significant computational burden or be prone to ambiguous mismatches. In this paper, we demonstrate that incorporating simple pseudo-depth and confidence matrices into the similarity cost can compensate for the missing information of powerful association cues during occlusion, thereby enhancing the effectiveness of the association results. Additionally, based on the assumption of virtual curved trajectories and velocity direction during congestion, our method, Enhanced Similarity Costs SORT (ESC-SORT), achieves performance comparable to state-of-the-art trackers on the DanceTrack, MOT20 and MOT17 benchmarks. With superior performance, our tracker generalizes to various real-world scenarios in a plug-and-play manner while remaining simple, online, and real-time. The code and models are available at https://gitee.com/SylviaCL/esc-sort.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.