Fine-grained classification on Flower102 (test)
97.8AccuracyLinear SVM
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Linear SVMfeature type=Global features, ground-truth part annotation=None2019.09 | 97.8 | — | — | — | |
| Cui et al.feature type=Global features, ground-truth part annotation=None2019.09 | 97.7 | — | — | — | |
| MC-LossBase Model=B-CNN2020.02 | 97.7 | — | — | — | |
| Our parts (classification-specific part estimation)feature type=Part-based features, maximum # of parts=4, ground-truth part annotation=None2019.09 | 96.9 | — | — | — | |
| MC-LossBase Model=ResNet502020.02 | 96.8 | — | — | — | |
| Simon et al.feature type=Global features, pooling=alpha-pooling, ground-truth part annotation=None2019.09 | 96.7 | — | — | — | |
| MC-LossBase Model=VGG162020.02 | 96.1 | — | — | — | |
| Selective joint FTBase Model=ResNet1522020.02 | 95.8 | — | — | — | |
| Simon et al.feature type=Part-based features, maximum # of parts=20, ground-truth part annotation=None2019.09 | 95.3 | — | — | — | |
| PCBase Model=B-CNN2020.02 | 93.7 | — | — | — | |
| PC-BilinearCNNBackbone=Bilinear CNN, Pairwise Confusion (PC)=true2017.05 | 93.65 | — | 1.13 | — | |
| PC-ResNet-50Backbone=ResNet-50, Pairwise Confusion (PC)=true2017.05 | 93.5 | — | 1.04 | — | |
| BilinearCNNBackbone=Bilinear CNN2017.05 | 92.52 | — | — | — | |
| B-CNNBase Model=VGG162020.02 | 92.5 | — | — | — | |
| ResNet-50Backbone=ResNet-502017.05 | 92.46 | — | — | — | |
| PC-DenseNet-161Backbone=DenseNet-161, Pairwise Confusion (PC)=true2017.05 | 91.39 | — | 1.32 | — | |
| PCBase Model=DenseNet1612020.02 | 91.2 | — | — | — | |
| DenseNet-161Backbone=DenseNet-1612017.05 | 90.07 | — | — | — | |
| Overfeat2017.05 | 86.8 | — | — | — | |
| OverfeatBase Model=Overfeat2020.02 | 86.8 | — | — | — | |
| Det.+seg.Base Model=SVM2020.02 | 80.7 | — | — | — | |
| Det.+Seg.2017.05 | 80.66 | — | — | — | |
| ZS_EnEvaluation Protocol=Zero-shot2023.11 | 73.16 | — | — | — | |
| CLIPBackbone=ViT-B/16, Evaluation Protocol=Zero-shot2023.11 | 71.3 | — | — | — | |
| CLIPBackbone=ViT-B/32, Evaluation Protocol=Zero-shot2023.11 | 66.95 | — | — | — | |
| CLIPBackbone=RN50, Evaluation Protocol=Zero-shot2023.11 | 65.98 | — | — | — | |
| CLIPBackbone=RN101, Evaluation Protocol=Zero-shot2023.11 | 63.95 | — | — | — | |
| MTACategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 63.5 | — | — | 22.9 | |
| CLIPCategory=Baseline, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 63.2 | — | — | 0 | |
| MACCategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 62.8 | — | — | 45.6 | |
| R-TPTCategory=Tuning-based, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 62.6 | — | — | 39 | |
| CLIPBackbone=CLIP-ResNet50, Epsilon=1.02026.06 | 61.7 | — | — | 0 | |
| TPTBackbone=CLIP-ResNet50, Epsilon=1.0, Augmentation views=152026.06 | 61.4 | — | — | 0 | |
| TTCCategory=Tuning-free, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 61.1 | — | — | 4 | |
| TAPTCategory=Tuning-based, Backbone=CLIP-ViT-B/32, Attack Protocol=PGD-100 (ϵ = 4.0)2026.06 | 60.3 | — | — | 9 | |
| R-TPTBackbone=CLIP-ResNet50, Epsilon=1.0, Augmentation views=152026.06 | 58.8 | — | — | 43.5 | |
| SS-TPTBackbone=CLIP-ResNet50, Epsilon=1.0, Augmentation views=152026.06 | 58.8 | — | — | 50.6 | |
| EnsembleBackbone=CLIP-ResNet50, Epsilon=1.0, Augmentation views=152026.06 | 58.1 | — | — | 37.4 | |
| DOCBackbone=CLIP-ResNet50, Epsilon=1.0, Effective batch size=12026.06 | 56.1 | — | — | 7.2 | |
| TTCBackbone=CLIP-ResNet50, Epsilon=1.0, Effective batch size=12026.06 | 55.8 | — | — | 7.5 |