Generalized Category Discovery on Herbarium19
50.3Score (All Categories)SimGCD-CLIP + AL-GCD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SimGCD-CLIP + AL-GCDBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 50.3 | 63.6 | 43.1 | |
| GETBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 49.7 | 64.5 | 41.7 | |
| Sharpness-aware Dynamic Anchor Selection (Ours)Backbone=ViT-B/162025.12 | 49.6 | 61.9 | 43 | |
| InfoSculpt2026.01 | 48.5 | 61.2 | 36 | |
| SimGCD-CLIPBackbone=CLIP, Ground-truth number of categories K=Given, reproduced=true2026.03 | 47.9 | 62.3 | 39.9 | |
| SelEx-CLIP + AL-GCDBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 47.7 | 59 | 41.6 | |
| RLCDBackbone=DINO, Ground-truth number of categories K=Given2026.03 | 46.4 | 61.2 | 38.4 | |
| μGCDPre-training=DINO [14]2023.11 | 45.8 | 61.9 | 37.2 | |
| µGCDBackbone=ViT-B/162025.12 | 45.8 | 61.9 | 37.2 | |
| MTMC2026.01 | 45.6 | 57.8 | 39 | |
| CPTBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 45.3 | 61.2 | 36.7 | |
| LegoGCD2025.01 | 45.1 | 57.4 | 38.4 | |
| LegoGCD2026.01 | 45.1 | 57.4 | 38.4 | |
| LegoGCDBackbone=ViT-B/162025.12 | 45.1 | 57.4 | 38.4 | |
| SimGCD*variant=re-implemented2025.01 | 44.9 | 56.9 | 38.4 | |
| ProtoGCD2026.01 | 44.5 | 59.4 | 36.5 | |
| SelEx-CLIPBackbone=CLIP, Ground-truth number of categories K=Given, reproduced=true2026.03 | 44.4 | 57.1 | 37.6 | |
| SimGCD2025.01 | 44 | 58 | 36.4 | |
| SimGCD2026.01 | 44 | 58 | 36.4 | |
| SimGCDBackbone=ViT-B/162025.12 | 44 | 58 | 36.4 | |
| SimGCDBackbone=DINO, Ground-truth number of categories K=Given2026.03 | 44 | 58 | 36.4 | |
| PIM + L-RegCategory=InfoMax based2024.10 | 43.7 | 55.8 | 37.2 | |
| SPTNetBackbone=ViT-B/162025.12 | 43.4 | 58.7 | 35.2 | |
| SimGCDPre-training=DINO [14]2023.11 | 43.3 | 57.9 | 35.3 | |
| MIBPre-training=DINO [14]2023.11 | 42.3 | 56.1 | 34.8 | |
| PIMCategory=InfoMax based2024.10 | 42.3 | 56.1 | 34.8 | |
| CMS-CLIP + AL-GCDBackbone=CLIP, K (Ground-truth number of categories)=Not given2026.03 | 41.2 | 59 | 31.6 | |
| CMS-CLIP + AL-GCDBackbone=CLIP, Ground-truth number of categories K=Not Given2026.03 | 41.2 | 59 | 31.6 | |
| InfoSieveBackbone=ViT-B/162025.12 | 41 | 55.4 | 33.2 | |
| RIMCategory=InfoMax based2024.10 | 40.1 | 57.6 | 30.7 | |
| TIMCategory=InfoMax based2024.10 | 40.1 | 57.4 | 30.7 | |
| GCD-CLIPBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 39.7 | 58 | 29.9 | |
| SelExBackbone=DINO, Ground-truth number of categories K=Given2026.03 | 39.6 | 54.9 | 31.3 | |
| CMS-CLIP + AL-GCDBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 39.6 | 59.2 | 29 | |
| OpenConPre-training=DINO [14]2023.11 | 39.3 | 58.9 | 28.6 | |
| CMS-CLIP (CVPR 24)Backbone=CLIP, K (Ground-truth number of categories)=Not given2026.03 | 38.8 | 57.7 | 28.6 | |
| CMS-CLIPBackbone=CLIP, Ground-truth number of categories K=Not Given2026.03 | 38.8 | 57.7 | 28.6 | |
| CMS (CVPR 24)Backbone=DINO, K (Ground-truth number of categories)=Not given2026.03 | 37.4 | 56.5 | 27.1 | |
| CMSBackbone=DINO, Ground-truth number of categories K=Not Given2026.03 | 37.4 | 56.5 | 27.1 | |
| GCD (CVPR 22)Backbone=DINO, K (Ground-truth number of categories)=Not given2026.03 | 37.2 | 51.7 | 29.4 | |
| GCDBackbone=DINO, Ground-truth number of categories K=Not Given2026.03 | 37.2 | 51.7 | 29.4 | |
| PromptCALPre-training=DINO [14]2023.11 | 37 | 52 | 28.9 | |
| PromptCALBackbone=ViT-B/162025.12 | 37 | 52 | 28.9 | |
| GPC (ICCV 23)Backbone=DINO, K (Ground-truth number of categories)=Not given2026.03 | 36.5 | 51.7 | 27.9 | |
| GPCBackbone=DINO, Ground-truth number of categories K=Not Given2026.03 | 36.5 | 51.7 | 27.9 | |
| CMSBackbone=ViT-B/162025.12 | 36.4 | 54.9 | 26.4 | |
| CMSBackbone=DINO, Ground-truth number of categories K=Given2026.03 | 36.4 | 54.9 | 26.4 | |
| CMS-CLIPBackbone=CLIP, Ground-truth number of categories K=Given2026.03 | 36.2 | 56.5 | 25.3 | |
| GCDPre-training=DINO [14]2023.11 | 35.4 | 51 | 27 | |
| GCD2024.10 | 35.4 | 51 | 27 | |
| GCD2025.01 | 35.4 | 51 | 27 | |
| GCD2026.01 | 35.4 | 51 | 27 | |
| GCDBackbone=ViT-B/162025.12 | 35.4 | 51 | 27 | |
| GCDBackbone=DINO, Ground-truth number of categories K=Given2026.03 | 35.4 | 51 | 27 | |
| UNO+Pre-training=DINO [14]2023.11 | 28.3 | 53.7 | 12.8 | |
| UNO+2024.10 | 28.3 | 53.7 | 14.7 | |
| UNO+2025.01 | 28.3 | 53.7 | 14.7 | |
| UNO+Backbone=ViT-B/162025.12 | 28.3 | 53.7 | 14.7 | |
| RankStats+Pre-training=DINO [14]2023.11 | 27.9 | 55.8 | 12.8 | |
| RankStats+2024.10 | 27.9 | 55.8 | 12.8 | |
| RS+2025.01 | 27.9 | 55.8 | 12.8 | |
| RS+Backbone=ViT-B/162025.12 | 27.9 | 55.8 | 12.8 | |
| ORCABackbone=ViT-B/162024.10 | 25 | 30.6 | 19.8 | |
| ORCAVariant=Original2024.10 | 22.9 | 25.9 | 21.3 | |
| ORCAPre-training=DINO [14]2023.11 | 20.9 | 30.9 | 15.5 | |
| ORCA2025.01 | 20.9 | 30.9 | 15.5 | |
| ORCABackbone=ViT-B/162025.12 | 20.9 | 30.9 | 15.5 | |
| k-meansPre-training=DINO [14]2023.11 | 13 | 12.2 | 13.4 | |
| k-means2025.01 | 13 | 12.2 | 13.4 | |
| k-meansBackbone=ViT-B/162025.12 | 13 | 12.2 | 13.4 | |
| K-means2024.10 | 12.9 | 12.9 | 12.8 |