Models & Raw results:
- Google Drive
- baidu yun, code: rcsn
VOS test configuration directory: _experiments/sat/test/
| MODEL | Pipeline | Dataset | J&F-Mean | J-Mean | J-Recall | J-Decay | F-Mean | F-Recall | F-Decaly | FPS@GTX2080Ti | Config. Filename |
|---|---|---|---|---|---|---|---|---|---|---|---|
| SAT-Res50 | StateAwareTracker | DAVIS2017_val | 0.712 | 0.676 | 0.781 | 0.144 | 0.748 | 0.854 | 0.18 | ~35 | sat_res50-davis17.yaml |
| SAT-Res18 | StateAwareTracker | DAVIS2017_val | 0.696 | 0.665 | 0.761 | 0.162 | 0.727 | 0.820 | 0.191 | ~50 | sat_res18-davis17.yaml |
| SAT-Res50 | StateAwareTracker | DAVIS2016_val | 0.821 | 0.818 | 0.938 | 0.022 | 0.824 | 0.931 | 0.05 | ~35 | sat_res50-davis16.yaml |
Nota:
[1] We reimplement SAT with pytorch. The J&F is slightly lower than what we trained with internal framework. We will continue refining the training code.