Short name:
CSTrack
Detector:
Private
Description:
CSTrack proposes a strong ReID based one-shot MOT framework. It includes a novel cross-correlation network that can effectively impel the separate branches to learn task-dependent representations, and a scale-aware attention network that learns discriminative embeddings to improve the ReID capability. This work also provides an analysis of the weak data association ability in one-shot MOT methods. Our improvements make the data association ability of our one-shot model is comparable to two-stage methods while running more faster.
Reference:
C. Liang, Z. Zhang, Y. Lu, X. Zhou, B. Li. Rethinking the competition between detection and ReID in Multi-Object Tracking. In arXiv:2010.12138 [cs], 2020.
Last submitted:
December 23, 2020 (3 years ago)
Published:
December 15, 2020 at 07:25:46 CET
Submissions:
2
Project page / code:
Open source:
Yes
Hardware:
RTX2080ti
Runtime:
4.5 Hz
Benchmark performance:
Sequence | MOTA | IDF1 | HOTA | MT | ML | FP | FN | Rcll | Prcn | AssA | DetA | AssRe | AssPr | DetRe | DetPr | LocA | FAF | ID Sw. | Frag |
MOT20 | 66.6 | 68.6 | 54.0 | 626 (50.4) | 192 (15.5) | 25,404 | 144,358 | 72.1 | 93.6 | 54.0 | 54.2 | 57.6 | 77.8 | 58.5 | 76.0 | 81.5 | 5.7 | 3,196 (44.3) | 7,632 (105.9) |
Detailed performance:
Sequence | MOTA | IDF1 | HOTA | MT | ML | FP | FN | Rcll | Prcn | AssA | DetA | AssRe | AssPr | DetRe | DetPr | LocA | FAF | ID Sw. | Frag |
MOT20-04 | 80.2 | 79.0 | 61.2 | 439 | 32 | 12,899 | 40,226 | 85.3 | 94.8 | 59.3 | 63.4 | 62.9 | 79.0 | 68.6 | 76.2 | 81.2 | 6.2 | 1,021 | 3,061 |
MOT20-06 | 50.7 | 53.9 | 43.7 | 80 | 86 | 6,234 | 58,185 | 56.2 | 92.3 | 44.1 | 43.5 | 47.3 | 75.1 | 46.5 | 76.5 | 82.4 | 6.2 | 1,065 | 2,547 |
MOT20-07 | 73.5 | 66.4 | 55.0 | 73 | 3 | 2,475 | 5,910 | 82.1 | 91.7 | 50.0 | 61.3 | 54.2 | 76.1 | 67.9 | 75.8 | 82.4 | 4.2 | 398 | 523 |
MOT20-08 | 42.5 | 50.7 | 39.3 | 34 | 71 | 3,796 | 40,037 | 48.3 | 90.8 | 42.6 | 36.4 | 45.6 | 75.4 | 39.0 | 73.4 | 81.3 | 4.7 | 712 | 1,501 |
Raw data: