Image Classification on iNaturalist 18
87Overall AccuracyCategory Extrapolation (Ours)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Category Extrapolation (Ours)Pre-training paradigm=DINOv2, Data configuration=proposed method2024.10 | 87 | 86.4 | 87.4 | 86.7 | |
| Baseline + SDPre-training paradigm=DINOv2, Data configuration=selected neighbor categories2024.10 | 85.9 | 85.8 | 86.5 | 85 | |
| BaselinePre-training paradigm=DINOv2, Data configuration=standard2024.10 | 85 | 85.7 | 86.2 | 84.2 | |
| dBOT-RefinedBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 84.5 | — | — | — | |
| dBOTBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 84.1 | — | — | — | |
| MAE-RefinedBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 83.3 | — | — | — | |
| MAE-CTBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 82.9 | — | — | — | |
| MAEBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 82.7 | — | — | — | |
| dBOT-RefinedBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 82.1 | — | — | — | |
| dBOTBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 81.9 | — | — | — | |
| D2V2-RefinedBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 81.6 | — | — | — | |
| D2V2Backbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 81 | — | — | — | |
| MAE-CTBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 80.9 | — | — | — | |
| MAE-RefinedBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 80.9 | — | — | — | |
| MAEBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 80.7 | — | — | — | |
| D2V2-RefinedBackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 79.8 | — | — | — | |
| D2V2Backbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 79.6 | — | — | — | |
| Category Extrapolation (Ours)Pre-training paradigm=CLIP, Data configuration=proposed method2024.10 | 78.5 | 79.5 | 79.3 | 77.3 | |
| CrossMAE-RefinedBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 78.4 | — | — | — | |
| Category Extrapolation (Ours)Pre-training paradigm=Scratch, Data configuration=proposed method2024.10 | 78 | 78.9 | 78.2 | 77.5 | |
| CrossMAEBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 77.7 | — | — | — | |
| iBOTBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 76.9 | — | — | — | |
| MugsBackbone=ViT-L/16, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 76.9 | — | — | — | |
| Baseline + SDPre-training paradigm=Scratch, Data configuration=selected neighbor categories2024.10 | 76.8 | 78.6 | 77.2 | 75.7 | |
| Baseline + SDPre-training paradigm=CLIP, Data configuration=selected neighbor categories2024.10 | 76.7 | 78.5 | 77.9 | 74.6 | |
| MAE-RefBackbone=ViT-2B/14, Protocol=10-shot Fine-tuning2024.02 | 76.5 | — | — | — | |
| BaselinePre-training paradigm=Scratch, Data configuration=standard2024.10 | 76.1 | 78.5 | 76.9 | 74.6 | |
| I-JEPABackbone=ViT-H/14, Evaluation Protocol=Full fine-tuning (100% labels)2024.02 | 75.9 | — | — | — | |
| dBOT-RefBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 75.8 | — | — | — | |
| Baseline + RDPre-training paradigm=DINOv2, Data configuration=random auxiliary data2024.10 | 75.4 | 82.3 | 76.1 | 72.6 | |
| MAE-RefBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 75.4 | — | — | — | |
| BaselinePre-training paradigm=CLIP, Data configuration=standard2024.10 | 75 | 77.8 | 76.5 | 72.5 | |
| D2V2-RefinedBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 74.8 | — | — | — | |
| MAE-CTBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 74.7 | — | — | — | |
| D2V2Backbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 73.4 | — | — | — | |
| dBOTBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 73.2 | — | — | — | |
| DeiT-BaseParams (M)=86.5, FLOPs (G)=17.56, Structured Compression=false, Fine-tuning=true2025.05 | 73.2 | — | — | — | |
| AttnScale=Base, Params (M)=86.52026.05 | 73.2 | — | — | — | |
| dBOT-RefBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 73 | — | — | — | |
| MAEBackbone=ViT-2B/14, Protocol=10-shot Fine-tuning2024.02 | 72.2 | — | — | — | |
| MSNEvaluation Protocol=Fine-Tuning, Backbone=ViT-Base/16, Pre-training=ImageNet-1K2022.04 | 72.1 | — | — | — | |
| DINOEvaluation Protocol=Fine-Tuning, Backbone=ViT-Base/16, Pre-training=ImageNet-1K2022.04 | 72 | — | — | — | |
| D2V2-RefinedBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 71.8 | — | — | — | |
| MAE-RefBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 71.7 | — | — | — | |
| MAEBackbone=ViT-H/14, Protocol=10-shot Fine-tuning2024.02 | 71.7 | — | — | — | |
| dBOTBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 71.2 | — | — | — | |
| CAEBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 71.2 | — | — | — | |
| BiMambaScale=Base, Params (M)=89.92026.05 | 71.2 | — | — | — | |
| D2V2Backbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 70.6 | — | — | — | |
| MAE-RefinerBackbone=ViT-2B/14, Protocol=Linear Probe2024.02 | 69.6 | — | — | — | |
| MAE-RefBackbone=ViT-2B/14, Protocol=Linear Probe2024.02 | 69.6 | — | — | — | |
| MAEBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 68.9 | — | — | — | |
| Baseline + RDPre-training paradigm=Scratch, Data configuration=random auxiliary data2024.10 | 68.4 | 76.4 | 70.1 | 64.3 | |
| CrossMAE-RefBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 67.8 | — | — | — | |
| Baseline + RDPre-training paradigm=CLIP, Data configuration=random auxiliary data2024.10 | 67.7 | 75.1 | 69.8 | 63.1 | |
| BiMambaScale=Small, Params (M)=26.92026.05 | 67.1 | — | — | — | |
| FAR-SmallParams (M)=23.9, FLOPs (G)=4.82, Structured Compression=false, Fine-tuning=true2025.05 | 67 | — | — | — | |
| FAR-BaseParams (M)=83.2, FLOPs (G)=17.31, Structured Compression=false, Fine-tuning=true2025.05 | 67 | — | — | — | |
| BiLSTMScale=Small, Params (M)=23.92026.05 | 67 | — | — | — | |
| BiLSTMScale=Base, Params (M)=83.22026.05 | 67 | — | — | — | |
| MugsBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 66.9 | — | — | — | |
| DeiT-SmallParams (M)=22.1, FLOPs (G)=4.60, Structured Compression=false, Fine-tuning=true2025.05 | 66.8 | — | — | — | |
| AttnScale=Small, Params (M)=22.92026.05 | 66.8 | — | — | — | |
| FAR-Base (Pruned)Params (M)=73.0, FLOPs (G)=14.80, Structured Compression=true, Fine-tuning=true2025.05 | 66 | — | — | — | |
| dBOT-RefBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 65.8 | — | — | — | |
| iBOTBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 65.5 | — | — | — | |
| FAR-Small (Pruned)Params (M)=20.3, FLOPs (G)=3.90, Structured Compression=true, Fine-tuning=true2025.05 | 65.4 | — | — | — | |
| BiMambaScale=Tiny, Params (M)=7.72026.05 | 64.9 | — | — | — | |
| FAR-TinyParams (M)=7.5, FLOPs (G)=1.45, Structured Compression=false, Fine-tuning=true2025.05 | 64.7 | — | — | — | |
| BiLSTMScale=Tiny, Params (M)=7.72026.05 | 64.7 | — | — | — | |
| MAE-RefinerBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 64.6 | — | — | — | |
| MAE-RefBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 64.6 | — | — | — | |
| CrossMAEBackbone=ViT-L/16, Protocol=10-shot Fine-tuning2024.02 | 63.7 | — | — | — | |
| MAE-RefBackbone=ViT-2B/14, Protocol=5-shot Fine-tuning2024.02 | 63.5 | — | — | — | |
| FAR-Tiny (Pruned)Params (M)=5.7, FLOPs (G)=1.08, Structured Compression=true, Fine-tuning=true2025.05 | 62.9 | — | — | — | |
| MAE-CTBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 62.8 | — | — | — | |
| MAE-CTBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 62.8 | — | — | — | |
| MAE-RefBackbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 62.4 | — | — | — | |
| DeiT-TinyParams (M)=5.7, FLOPs (G)=1.25, Structured Compression=false, Fine-tuning=true2025.05 | 62.4 | — | — | — | |
| AttnScale=Tiny, Params (M)=5.72026.05 | 62.4 | — | — | — | |
| dBOT-RefBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 61.7 | — | — | — | |
| MugsBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 61.5 | — | — | — | |
| MugsBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 61.5 | — | — | — | |
| dBOT-RefBackbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 60.9 | — | — | — | |
| MAE-RefinerBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 60.6 | — | — | — | |
| MAE-RefBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 60.6 | — | — | — | |
| CrossMAE-RefBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 60.5 | — | — | — | |
| MAE-CTBackbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 60.1 | — | — | — | |
| D2V2-RefinedBackbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 59.2 | — | — | — | |
| dBOT-RefBackbone=ViT-L/16, Protocol=5-shot Fine-tuning2024.02 | 58.8 | — | — | — | |
| MAE-RefBackbone=ViT-L/16, Protocol=5-shot Fine-tuning2024.02 | 58 | — | — | — | |
| D2V2-RefinedBackbone=ViT-L/16, Protocol=5-shot Fine-tuning2024.02 | 56.3 | — | — | — | |
| iBOTBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 56 | — | — | — | |
| iBOTBackbone=ViT-L/16, Protocol=Linear Probe2024.02 | 56 | — | — | — | |
| D2V2Backbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 55.7 | — | — | — | |
| TC-JEPAViT=H/14, Protocol=linear probing2026.05 | 54.8 | — | — | — | |
| CAEBackbone=ViT-L/16, Protocol=5-shot Fine-tuning2024.02 | 54.5 | — | — | — | |
| D2V2-RefinedBackbone=ViT-H/14, Protocol=Linear Probe2024.02 | 54.4 | — | — | — | |
| MAEBackbone=ViT-2B/14, Protocol=5-shot Fine-tuning2024.02 | 53.7 | — | — | — | |
| MAEBackbone=ViT-H/14, Protocol=5-shot Fine-tuning2024.02 | 53.3 | — | — | — |