IAMOT: Multi-object tracking with integrated heads and attention mechanism

Abstract

Multi-object tracking is one of the most fundamental problems in computer vision with wide industrial applications. It involves the association of multiple targets across consecutive frames. The current trend has been gradually switched from tracking-by-detection paradigm to one-shot trackers. While special efforts have been made on anchor-free methods which turned out to be more suitable for the extraction of re-ID features (than anchor-based methods). However, most of the works used center points to represent the targets and regressed other properties, which lost global information and were unfavorable to ID switches. To address the above issues, we propose IAMOT, a simple yet effective network based on anchor-free architecture. Our network predicts the center point and derives the corner points of an object to obtain a refined bounding box through integrated heads. It also utilizes an additional attention module to weaken the occurrence of ID switches. Extensive experiments have been conducted on 4 mainstream datasets, where IAMOT exhibits superior performances by surpassing other state-of-the-art methods. Especially, we achieve 60.6% MOTA on MOT15, 75.8% MOTA on MOT16, 74.4% MOTA on MOT17 and 64.1% MOTA on MOT20, respectively.