Rethinking the Competition Between Detection and ReID in Multiobject Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• One-shot multi-object tracking (MOT) models jointly learn detection and identification embeddings for efficiency.
• Existing one-shot MOT methods often overlook the inherent differences and relations between detection and re-identification (ReID), treating them as isolated tasks.
• This isolation leads to suboptimal performance compared to two-stage tracking methods.

Purpose of the Study:

• To address the detrimental task competition in one-shot MOT by improving the learning of task-dependent representations.
• To enhance the cooperation between detection and re-identification (ReID) tasks within a unified framework.
• To improve the association capability of identity (ID) embeddings through scale-aware attention.

Main Methods:

• Proposing a novel reciprocal network (REN) with self-relation and cross-relation mechanisms to mitigate task competition.
• Introducing a scale-aware attention network (SAAN) to prevent semantic misalignment and improve ID embedding association.
• Integrating REN and SAAN into a one-shot online MOT system to create the CSTrack tracker.

Main Results:

• CSTrack achieves state-of-the-art performance on the MOT16, MOT17, and MOT20 benchmark datasets.
• The proposed tracker demonstrates efficiency, running at 16.4 FPS on a single GPU, with a lightweight version reaching 34.6 FPS.
• The method effectively alleviates task competition and improves cooperation between detection and ReID.

Conclusions:

• The CSTrack tracker offers a strong and efficient solution for one-shot online multi-object tracking.
• The reciprocal network and scale-aware attention mechanisms are key to achieving superior performance.
• The findings highlight the importance of considering task interdependencies in MOT model design.