Spatial-Attention Location-Aware Multi-Object Tracking

Abstract

Most existing one-shot multi-object tracking (MOT) methods have already made great progress in jointly accomplishing detection and re-identification tasks with a single network. However, they often ignore detection misalignments and may heavily rely on tightly enclosed image patches, which aggravates the dependence of tracking and re-identification performance on detection accuracy. We evaluate the confidence of appearance embeddings with predicted location precision, which alleviates this heavy dependence. To deal with misalignments, person search is introduced to jointly train the detection and re-identification using proposals via multi-task learning. In addition, we equip our network with feature fusion strategy at different scales and spatial-channel attention module to narrow the semantic gaps and focus on informative regions. The designed network serves as an online multi-object tracker and can be easily trained end-to-end. Extensive experiments show that our proposed method achieves the competitive performance against most state-of-the-art methods on several MOTChallenge benchmarks while running at over 12 FPS.