Action Recognition on SSv2 10K clips (test)
20.4Probe AccuracyTemporal+M SAE
Evaluation Results
| Method | Links | |
|---|---|---|
| Temporal+M SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Temporal+Matryoshka2026.04 | 20.4 | |
| Temporal SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Temporal2026.04 | 20.2 | |
| Separate SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Separate2026.04 | 20.2 | |
| Separate+M SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Separate+Matryoshka2026.04 | 19.7 | |
| Standard SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Standard2026.04 | 19.1 | |
| Temporal+M SAE (Coherence-optimized)Backbone=DINOv2, Training Scale=10K clips, Variant=Temporal+Matryoshka, tau=0.2, optimization=Coherence-optimized2026.04 | 19.1 | |
| Raster SAEBackbone=DINOv2, Training Scale=10K clips, Variant=Raster2026.04 | 18.3 | |
| Raw DINOv2Backbone=DINOv2, Evaluation Protocol=Linear Probing2026.04 | 17.4 |