Temporal-invariant Feature Extraction on DAVIS
0.6< δ^0 ErrorDINOv3-L
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DINOv3-LBackbone=DINOv3-L2026.05 | 0.6 | 6.2 | 40.7 | 13.7 | |
| DINOv3-L + C4GBackbone=DINOv3-L, Feature Enhancement=C4G2026.05 | 2.2 | 19.8 | 63.1 | 27.1 | |
| VGGT-TrackingBackbone=VGGT-Tracking2026.05 | 2.7 | 17.9 | 54 | 23.1 | |
| VGGT-Tracking + C4GBackbone=VGGT-Tracking, Feature Enhancement=C4G2026.05 | 4.4 | 23.4 | 68.7 | 30.5 |