Fine-grained classification on Stanford Cars
95AccuracyMMAL-Net
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MMAL-NetTrain Anno.=false2020.03 | 95 | — | |
| DCLTrain Anno.=false2020.03 | 94.5 | — | |
| DFN2B-CLIP-B/16Evaluation protocol=frozen-backbone linear probing, Crop size=224 x 2242026.05 | 94.33 | — | |
| NTS-NetTrain Anno.=false2020.03 | 93.9 | — | |
| DFL-CNNTrain Anno.=false2020.03 | 93.8 | — | |
| TASNTrain Anno.=false2020.03 | 93.8 | — | |
| HBPTrain Anno.=false2020.03 | 93.7 | — | |
| DINOv3-BEvaluation protocol=frozen-backbone linear probing, Crop size=256 x 2562026.05 | 93.32 | — | |
| OSME+MAMCTrain Anno.=false2020.03 | 93 | — | |
| PCTrain Anno.=false2020.03 | 92.9 | — | |
| S2-FracMixBackbone=ResNeXt502026.06 | 92.86 | — | |
| MA-CNNTrain Anno.=false2020.03 | 92.8 | — | |
| ResNet-50 (ReLabel-trained)Backbone=ResNet-50, Pre-training strategy=ReLabel2021.01 | 92.73 | — | |
| OpenCLIP-B/16Evaluation protocol=frozen-backbone linear probing, Crop size=224 x 2242026.05 | 92.65 | — | |
| ResNet-50 (Baseline)Backbone=ResNet-50, Pre-training strategy=Standard2021.01 | 92.64 | — | |
| SigLIP2-B/16-256Evaluation protocol=frozen-backbone linear probing, Crop size=256 x 2562026.05 | 92.55 | — | |
| RA-CNNTrain Anno.=false2020.03 | 92.5 | — | |
| KPTrain Anno.=false2020.03 | 92.4 | — | |
| VECA-BEvaluation protocol=frozen-backbone linear probing, Crop size=256 x 256, Core budget (C)=642026.05 | 91.84 | — | |
| AdAutoMixBackbone=ResNeXt502026.06 | 91.59 | — | |
| AutoMixBackbone=ResNeXt502026.06 | 91.38 | — | |
| ResizeMixBackbone=ResNeXt502026.06 | 91.36 | — | |
| Bilinear-CNNTrain Anno.=false2020.03 | 91.3 | — | |
| PuzzleMixBackbone=ResNeXt502026.06 | 91.29 | — | |
| CutMixBackbone=ResNeXt502026.06 | 91.22 | — | |
| FMixBackbone=ResNeXt502026.06 | 90.9 | — | |
| MixUpBackbone=ResNeXt502026.06 | 90.81 | — | |
| DINOv2-Reg-BEvaluation protocol=frozen-backbone linear probing, Crop size=224 x 2242026.05 | 90.62 | — | |
| SaliencyMixBackbone=ResNeXt502026.06 | 90.6 | — | |
| S2-FracMixBackbone=ResNet182026.06 | 90.56 | — | |
| ManifoldMixBackbone=ResNeXt502026.06 | 90.2 | — | |
| VECA-BEvaluation protocol=frozen-backbone linear probing, Crop size=256 x 256, Core budget (C)=82026.05 | 90.18 | — | |
| VanillaBackbone=ResNeXt502026.06 | 90.15 | — | |
| AdAutoMixBackbone=ResNet182026.06 | 89.19 | — | |
| C-RADIOv3-BEvaluation protocol=frozen-backbone linear probing, Crop size=256 x 2562026.05 | 89.14 | — | |
| AutoMixBackbone=ResNet182026.06 | 88.89 | — | |
| ResizeMixBackbone=ResNet182026.06 | 88.17 | — | |
| PuzzleMixBackbone=ResNet182026.06 | 87.78 | — | |
| FMixBackbone=ResNet182026.06 | 87.55 | — | |
| CutMixBackbone=ResNet182026.06 | 87.48 | — | |
| DINOv2-BEvaluation protocol=frozen-backbone linear probing, Crop size=224 x 2242026.05 | 86.62 | — | |
| SaliencyMixBackbone=ResNet182026.06 | 86.48 | — | |
| VanillaBackbone=ResNet182026.06 | 86.32 | — | |
| MixUpBackbone=ResNet182026.06 | 86.27 | — | |
| ManifoldMixBackbone=ResNet182026.06 | 85.88 | — | |
| CLIP-B/16Evaluation protocol=frozen-backbone linear probing, Crop size=224 x 2242026.05 | 85.14 | — | |
| HSFMTraining protocol=Pretrained2026.03 | 85.1 | — | |
| ERMTraining protocol=Standard ERM2026.03 | 83.98 | — | |
| HSFMTraining protocol=ERM2026.03 | 83.44 | — | |
| DFRTraining protocol=ERM2026.03 | 82.74 | — | |
| MTAPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 78.4 | 36.6 | |
| TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 78 | 0 | |
| C-TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 77.4 | 0 | |
| R-TPTPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 77.2 | 49.1 | |
| CLIPPre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 76.8 | 0 | |
| EnsemblePre-trained Backbone=CLIP-ViT-L/14, Perturbation budget (epsilon)=4.02025.04 | 76.3 | 40.5 | |
| MTABackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 67.7 | 18.5 | |
| AGCBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 67.3 | 93.2 | |
| R-TPTBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 67 | 34.7 | |
| EnsembleBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 66 | 41.8 | |
| CLIPBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 65.5 | 0 | |
| TTPBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 65.4 | 37.4 | |
| MTABackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 63.4 | 26.4 | |
| R-TPTPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 63.1 | 28.4 | |
| R-TPTBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 63.1 | 28.4 | |
| AGCBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 62.2 | 87.4 | |
| ADAPTBackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Transductive2025.08 | 61.46 | — | |
| DiffTPTBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 60.71 | — | |
| CLIPPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 60.1 | 0 | |
| CLIPBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 60.1 | 0 | |
| DMNBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 60.02 | — | |
| EnsembleBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 59.8 | 33.1 | |
| TTPBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 59.8 | 29.8 | |
| TPSBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 59.4 | — | |
| DPEBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 59.26 | — | |
| DotaBackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Online2025.08 | 58.72 | — | |
| TPTBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 58.46 | — | |
| ADAPTBackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Online2025.08 | 58.38 | — | |
| StatABackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Transductive2025.08 | 58.2 | — | |
| BCABackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Online2025.08 | 58.13 | — | |
| R-TPTPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 58.1 | 42.9 | |
| TransCLIPBackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Transductive2025.08 | 57.9 | — | |
| TDABackbone=CLIP-RN50, BP-free=true, Evaluation Protocol=Online2025.08 | 57.78 | — | |
| C-TPTBackbone=CLIP-RN50, BP-free=false, Evaluation Protocol=Online2025.08 | 56.5 | — | |
| CLIPPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 55.7 | 0 | |
| CLIPBackbone=CLIP-RN50, BP-free=N/A, Evaluation Protocol=Online2025.08 | 55.7 | — | |
| TTCBackbone=ViT-B/16, Epsilon (ε)=4/2552026.05 | 55 | 2.9 | |
| TTCBackbone=ViT-B/32, Epsilon (ε)=4/2552026.05 | 48.1 | 2.3 | |
| FAREPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 39.2 | 34.4 | |
| APT+TeCoAPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 33.9 | 33.6 | |
| DFRTraining protocol=Pretrained2026.03 | 32.75 | — | |
| TeCoAPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 22.4 | 22.3 | |
| APT+TeCoAPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 20.8 | 18.9 | |
| TeCoAPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 10.2 | 9.1 | |
| APTPre-trained Backbone=ResNet50, Perturbation budget (epsilon)=1.02025.04 | 8.5 | 0.6 | |
| APTPre-trained Backbone=ViT-B/32, Perturbation budget (epsilon)=4.02025.04 | 1.5 | 0.1 |