FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking

FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking

19 Oct 2021 | Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, Wenyu Liu
Multi-object tracking (MOT) is a significant challenge in computer vision with wide-ranging applications. Formulating MOT as a single network that jointly learns object detection and re-identification (re-ID) tasks is appealing due to its high computational efficiency. However, these two tasks often compete, leading to biased performance. Previous methods typically treat re-ID as a secondary task, which negatively impacts its accuracy. To address this issue, the authors propose FairMOT, an anchor-free object detection architecture based on CenterNet. FairMOT treats both detection and re-ID tasks equally, addressing three key issues: the use of anchors, feature sharing, and feature dimensionality. The approach achieves high accuracy for both detection and tracking, outperforming state-of-the-art methods on multiple public datasets. FairMOT ranks first among all trackers on the 2DMOT15, MOT16, MOT17, and MOT20 datasets. The method is simple, runs at 30 FPS on a single RTX 2080Ti GPU, and provides valuable insights into the relationship between detection and re-ID in MOT.Multi-object tracking (MOT) is a significant challenge in computer vision with wide-ranging applications. Formulating MOT as a single network that jointly learns object detection and re-identification (re-ID) tasks is appealing due to its high computational efficiency. However, these two tasks often compete, leading to biased performance. Previous methods typically treat re-ID as a secondary task, which negatively impacts its accuracy. To address this issue, the authors propose FairMOT, an anchor-free object detection architecture based on CenterNet. FairMOT treats both detection and re-ID tasks equally, addressing three key issues: the use of anchors, feature sharing, and feature dimensionality. The approach achieves high accuracy for both detection and tracking, outperforming state-of-the-art methods on multiple public datasets. FairMOT ranks first among all trackers on the 2DMOT15, MOT16, MOT17, and MOT20 datasets. The method is simple, runs at 30 FPS on a single RTX 2080Ti GPU, and provides valuable insights into the relationship between detection and re-ID in MOT.
Reach us at info@study.space
Understanding FairMOT%3A On the Fairness of Detection and Re-identification in Multiple Object Tracking