Single-label environmental sound classification on ESC-50
96.9AccuracyMJEPA ViT-g + data scaling
Evaluation Results
| Method | Links | |
|---|---|---|
| MJEPA ViT-g + data scalingEvaluation protocol=Frozen feature probing, Params=1B, Modality=A+V, Pre-train data=AS+VM2M2026.06 | 96.9 | |
| MJEPA ViT-L + data scalingEvaluation protocol=Frozen feature probing, Params=300M, Modality=A+V, Pre-train data=AS+VM2M2026.06 | 96.8 | |
| XKDEvaluation protocol=Full-model finetuning, Params=170M, Modality=A+V, Pre-train data=AS2026.06 | 96.5 | |
| AJEPAEvaluation protocol=Full-model finetuning, Params=85M, Modality=A, Pre-train data=AS2026.06 | 96.3 | |
| SSLAMEvaluation protocol=Full-model finetuning, Params=88M, Modality=A, Pre-train data=AS2026.06 | 96.2 | |
| EquiAVEvaluation protocol=Full-model finetuning, Params=170M, Modality=A+V, Pre-train data=AS2026.06 | 96 | |
| MJEPA ViT-LEvaluation protocol=Frozen feature probing, Params=300M, Modality=A+V, Pre-train data=AS2026.06 | 95.2 | |
| MAViLEvaluation protocol=Full-model finetuning, Params=170M, Modality=A+V, Pre-train data=IN+AS2026.06 | 94.4 | |
| XKDEvaluation protocol=Frozen feature probing, Params=170M, Modality=A+V, Pre-train data=AS2026.06 | 93.6 | |
| SSLAMEvaluation protocol=Frozen feature probing, Params=88M, Modality=A, Pre-train data=AS2026.06 | 93.2 | |
| EquiAVEvaluation protocol=Frozen feature probing, Params=170M, Modality=A+V, Pre-train data=AS2026.06 | 93.2 | |
| Dasheng-1.2BEvaluation protocol=Frozen feature probing, Params=1.2B, Modality=A, Pre-train data=272k hrs mix2026.06 | 92.2 | |
| MAViLEvaluation protocol=Frozen feature probing, Params=170M, Modality=A+V, Pre-train data=IN+AS2026.06 | 90.8 | |
| SPEAREvaluation protocol=Frozen feature probing, Params=600M, Modality=A, Pre-train data=197k hrs mix2026.06 | 89.4 | |
| CAV-MAE SyncEvaluation protocol=Frozen feature probing, Params=170M, Modality=A+V, Pre-train data=AS2026.06 | 89.2 | |
| CAV-MAEEvaluation protocol=Frozen feature probing, Params=170M, Modality=A+V, Pre-train data=IN+AS2026.06 | 77.5 | |
| AJEPAEvaluation protocol=Frozen feature probing, Params=85M, Modality=A, Pre-train data=AS2026.06 | 74.4 |