ColTrack
October 16, 2023 ยท View on GitHub
The code has been moved to bytedance/ColTrack.
ColTrack not only outperforms state-of-the-art methods on large-scale datasets under high frame rates but also achieves higher and more stable performance under low frame rates. This allows it to obtain a higher equivalent FPS by reducing the frame rate requirement.
ColTrack: Collaborative Tracking Learning for Frame-Rate-Insensitive Multi-Object Tracking
Yiheng Liu, Junta Wu, Yi Fu
News
- The code can be found in https://github.com/bytedance/ColTrack.
- (2023.09.22) The code is in the company's open source review process and will be available soon.
- (2023.07) Our paper is accepted by ICCV 2023!
Tracking performance
| Dataset | HOTA | MOTA | IDF1 |
|---|---|---|---|
| MOT17 | 61.0 | 78.8 | 73.9 |
| Dancetrack | 72.6 | 92.1 | 74.0 |
| Dancetrack(+val) | 75.3 | 92.2 | 77.3 |
Visualization results on MOT challenge test set
