CVMay 22, 2023

Bridging the Gap Between End-to-end and Non-End-to-end Multi-Object Tracking

arXiv:2305.12724v138 citationsHas Code
Originality Incremental advance
AI Analysis

This work improves multi-object tracking for applications like autonomous driving by offering a more efficient and effective end-to-end method, though it is incremental as it builds on existing paradigms.

The paper tackles the performance gap between end-to-end and non-end-to-end multi-object tracking by addressing unbalanced training due to label assignment, proposing Co-MOT with a coopetition label assignment and shadow concept. It achieves 69.4% HOTA on DanceTrack and 52.8% TETA on BDD100K, with 38% FLOPs and 1.4x faster inference speed compared to MOTRv2.

Existing end-to-end Multi-Object Tracking (e2e-MOT) methods have not surpassed non-end-to-end tracking-by-detection methods. One potential reason is its label assignment strategy during training that consistently binds the tracked objects with tracking queries and then assigns the few newborns to detection queries. With one-to-one bipartite matching, such an assignment will yield unbalanced training, i.e., scarce positive samples for detection queries, especially for an enclosed scene, as the majority of the newborns come on stage at the beginning of videos. Thus, e2e-MOT will be easier to yield a tracking terminal without renewal or re-initialization, compared to other tracking-by-detection methods. To alleviate this problem, we present Co-MOT, a simple and effective method to facilitate e2e-MOT by a novel coopetition label assignment with a shadow concept. Specifically, we add tracked objects to the matching targets for detection queries when performing the label assignment for training the intermediate decoders. For query initialization, we expand each query by a set of shadow counterparts with limited disturbance to itself. With extensive ablations, Co-MOT achieves superior performance without extra costs, e.g., 69.4% HOTA on DanceTrack and 52.8% TETA on BDD100K. Impressively, Co-MOT only requires 38\% FLOPs of MOTRv2 to attain a similar performance, resulting in the 1.4$\times$ faster inference speed.

Code Implementations2 repos
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes