Arrow of Time (Directional Classification) on Something-Something V2 (val)
89.36AccuracyV-JEPA 2
Evaluation results
| Method | Links | |
|---|---|---|
| V-JEPA 2Modality=256×256 Video, Size=ViT-L, Pre-training Dataset=Real-World, Training Samples=22M (88×), Video Hours Eq.=≈1.73M (12, 357×), Pre-training explicitly includes SSv2=true, Evaluation Protocol=Linear Probing2026.05 | 89.36 | |
| TIMEModality=32×32 Tracks, Size=ViT-B, Pre-training Dataset=Synthetic, Training Samples=0.25M (1×), Video Hours Eq.=140 (1×), Pre-training explicitly includes SSv2=false, Evaluation Protocol=Linear Probing2026.05 | 88.53 | |
| RVMModality=224×224 Video, Size=ViT-B, Pre-training Dataset=Real-World, Training Samples=5.68M (23×), Video Hours Eq.=≈0.25M (1, 785×), Pre-training explicitly includes SSv2=true, Evaluation Protocol=Linear Probing2026.05 | 81.88 | |
| VideoMAEv2Modality=224×224 Video, Size=ViT-B, Pre-training Dataset=Real-World, Training Samples=5.36M (27×), Video Hours Eq.=≈7,500 (54×), Pre-training explicitly includes SSv2=true, Evaluation Protocol=Linear Probing2026.05 | 71.87 |