Play behavior classification on Poultry Play Recognition (LOCO folds)
77Macro F1PlayClass (Feat. + V-JEPA 2.1)
Evaluation results
| Method | Links | |
|---|---|---|
| PlayClass (Feat. + V-JEPA 2.1)K=32, Model Scale=ViT-L, Classifier=1D-CNN2026.05 | 77 | |
| PlayClass (Feat. + V-JEPA 2.1)K=32, Model Scale=ViT-B, Classifier=1D-CNN2026.05 | 76.5 | |
| V-JEPA 2.1K=32, Model Scale=ViT-L, Classifier=1D-CNN2026.05 | 76.3 | |
| V-JEPA 2.1K=32, Model Scale=ViT-B, Classifier=1D-CNN2026.05 | 75.8 | |
| V-JEPA 2K=32, Model Scale=ViT-L, Classifier=1D-CNN2026.05 | 74.3 | |
| VideoPrismK=8/16, Model Scale=ViT-L, Classifier=1D-CNN2026.05 | 74.1 | |
| DINOv3K=24, Model Scale=ViT-L, Classifier=1D-CNN2026.05 | 74 | |
| VideoPrismK=8/16, Model Scale=ViT-B, Classifier=1D-CNN2026.05 | 73.8 | |
| Features (MLP)K=—, Classifier=MLP2026.05 | 73.4 | |
| DINOv3K=24, Model Scale=ViT-B, Classifier=1D-CNN2026.05 | 70.7 | |
| PlayClass (Ablation: MLP classifier)K=—, Model Scale=ViT-L, Classifier=MLP2026.05 | -0.9 | |
| PlayClass (Ablation: no class weights)K=32, Model Scale=ViT-L, Training=minus class weights2026.05 | -1.5 | |
| PlayClass (Ablation: no label smoothing)K=32, Model Scale=ViT-L, Training=minus label smoothing2026.05 | -1.5 | |
| PlayClass (Ablation: 1 epoch)K=32, Model Scale=ViT-L, Training=1 epoch2026.05 | -4.8 |