Clustering on Pets
93.7NMIPRCut*
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| PRCut*Embedding Model=DINOv3-B, K=37, Q=0.772025.11 | 93.7 | 93.7 | — | 279 | 7.2 | |
| H-RCutEmbedding Model=DINOv3-B, K=37, Q=0.772025.11 | 93.4 | 93.3 | — | 276 | 7.1 | |
| H-NCutEmbedding Model=DINOv3-B, K=37, Q=0.772025.11 | 93.3 | 93.2 | — | 276 | 7.2 | |
| Spectral ClusteringEmbedding Model=DINOv3-B, K=37, Q=0.772025.11 | 91.2 | 89.8 | — | 256 | 6.7 | |
| CLIP (zero-shot)Backbone=ViT-B/32, Evaluation Protocol=Zero-shot with GT labels2026.04 | 87.6 | 84.9 | 75.4 | — | — | |
| Neural Tangent Kernel Spectral Clustering2026.02 | 84.9 | 72 | 64 | — | — | |
| KEC_TACBackbone=ViT-B/32, Training Strategy=TAC2026.04 | 84.8 | 79.7 | 71.5 | — | — | |
| TACBackbone=ViT-B/32, Training Strategy=TAC2026.04 | 83.6 | 77.9 | 69.1 | — | — | |
| GradNorm2026.02 | 81.5 | 72 | 62.8 | — | — | |
| GradNorm2025.10 | 81.5 | 72 | 62.8 | — | — | |
| KEC (no train)Backbone=ViT-B/32, Training Strategy=Training-free2026.04 | 81.2 | 67.8 | 63.3 | — | — | |
| KEC_TURTLEBackbone=ViT-B/32, Training Strategy=TURTLE2026.04 | 81 | 72.4 | 63.3 | — | — | |
| TAC (KMeans)2026.02 | 80.4 | 66.9 | 59.2 | — | — | |
| TAC2025.10 | 80.4 | 66.9 | 59.2 | — | — | |
| TAC (SC)2026.02 | 79.9 | 66 | 58.5 | — | — | |
| TAC (no train)Backbone=ViT-B/32, Training Strategy=Training-free2026.04 | 79.2 | 65.8 | 57.8 | — | — | |
| TAC_TURTLEBackbone=ViT-B/32, Training Strategy=TURTLE2026.04 | 73.7 | 63.4 | 52.5 | — | — | |
| TURTLE (1-space)Backbone=ViT-B/322026.04 | 71.5 | 60.9 | 48.9 | — | — | |
| SICBackbone=ViT-B/322026.04 | 67.7 | 51.6 | 42.6 | — | — | |
| CLIP (k-means)Backbone=ViT-B/32, Training Strategy=Training-free2026.04 | 65 | 51.3 | 40.5 | — | — | |
| VLM Caption + ClusterLLM=GPT-4o2026.04 | 56.6 | 47.8 | 37.4 | — | — |