Triplet odd-one-out on ImageNet-HSJ triplet-with-context
57.9FM AccuracyViT
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ViTBackbone=ViT, Representation=human-aligned2026.04 | 57.9 | 58.5 | 67.2 | 16.1 | 14.8 | |
| DINOv2Backbone=DINOv2, Representation=human-aligned2026.04 | 57.7 | 59.8 | 68.7 | 19.2 | 14.9 | |
| SigLIPBackbone=SigLIP, Representation=human-aligned2026.04 | 57.1 | 59.5 | 68.5 | 20 | 15 | |
| DINOv2Backbone=DINOv2, Representation=original2026.04 | 53.6 | 59 | 65.7 | 22.4 | 11.2 | |
| SigLIPBackbone=SigLIP, Representation=original2026.04 | 47.1 | 58.2 | 68.2 | 44.9 | 17.4 | |
| ViTBackbone=ViT, Representation=original2026.04 | 38.6 | 46 | 53.9 | 39.6 | 17.1 |