Few-shot classification on CIFAR-FS (5-way, 1-shot and 5-shot)
91.35Accuracy (5-way 1-shot)Ours-PCA 512
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Ours-PCA 512Backbone=DINOv2-L2026.05 | 91.35 | 97.53 | |
| Ours-RawBackbone=DINOv2-L2026.05 | 90.92 | 97.66 | |
| Ours-PCA 256Backbone=DINOv2-L2026.05 | 90.71 | 97.14 | |
| Ours-PCA 128Backbone=DINOv2-L2026.05 | 89.59 | 96.45 | |
| PT-MAPBackbone=WRN-28-10, Learning setup=Transductive, External data=false, External labels=false2022.04 | 87.7 | 90.7 | |
| Ours-ICA 128Backbone=DINOv2-L2026.05 | 85.59 | 96.53 | |
| SYNTRANSBackbone=ViT-S2026.05 | 84.64 | 90.81 | |
| SemFewBackbone=Swin-T2026.05 | 84.34 | 89.11 | |
| P>M>FBackbone=ViT-base, Pre-training=IN1K, Learning setup=Semi-supervised, External data=true, External labels=false2022.04 | 84.3 | 92.2 | |
| SYNTRANSBackbone=ResNet-122026.05 | 82.58 | 89.42 | |
| SP-CLIPBackbone=Visformer-T2026.05 | 82.18 | 88.24 | |
| Ours-ICA 256Backbone=DINOv2-L2026.05 | 81.76 | 96.71 | |
| P>M>FBackbone=ViT-Small, Pre-training=IN1K, Learning setup=Semi-supervised, External data=true, External labels=false2022.04 | 81.1 | 92.5 | |
| LPE-CLIPBackbone=ResNet-122026.05 | 80.62 | 86.22 | |
| SMKDBackbone=ViT-S2026.05 | 80.08 | 90.63 | |
| SIBBackbone=WRN-28-10, Learning setup=Transductive, External data=false, External labels=false2022.04 | 80 | 85.3 | |
| ECKPNBackbone=ResNet-122021.06 | 79.2 | 91 | |
| SEGABackbone=ResNet-122026.05 | 78.45 | 86 | |
| SUNBackbone=ViT-S2026.05 | 78.37 | 88.84 | |
| DPGNBackbone=ResNet-122021.06 | 77.9 | 90.2 | |
| FewTUREBackbone=Swin-T2026.05 | 77.76 | 88.9 | |
| PLCMBackbone=ResNet12, Learning setup=Semi-supervised, External data=true, External labels=false2022.04 | 77.6 | 86.1 | |
| ECKPNBackbone=Conv-42021.06 | 77.5 | 89.1 | |
| Fine-tuningBackbone=WRN-28-10, Learning setup=Transductive, External data=false, External labels=false2022.04 | 76.6 | 85.8 | |
| Fine-tuningBackbone=WRN-28-102021.03 | 76.58 | 85.79 | |
| DPGNBackbone=Conv-42021.06 | 76.4 | 88.4 | |
| RFSBackbone=SEResNet-122026.05 | 75.6 | 88.2 | |
| H-OTBackbone=WRN28+RTloss2022.10 | 75.4 | 87.5 | |
| MMLBackbone=ResNet-122021.03 | 75.28 | 85.95 | |
| Ours-ICA 512Backbone=DINOv2-L2026.05 | 75.19 | 95.56 | |
| Free-LunchBackbone=WRN28+RTloss2022.10 | 74.9 | 86.2 | |
| H-OTBackbone=ResNet122022.10 | 74.5 | 87.1 | |
| 4S-FSLBackbone=ResNet-122026.05 | 74.5 | 88.76 | |
| PEAGBackbone=ResNet-122021.03 | 74.27 | 83.89 | |
| RFSBackbone=ResNet122022.10 | 73.9 | 86.9 | |
| Free-LunchBackbone=ResNet122022.10 | 73.8 | 86 | |
| P>M>FBackbone=RN50, Pre-training=IN1K, Learning setup=Semi-supervised, External data=true, External labels=false2022.04 | 73.7 | 84 | |
| ProtoNetBackbone=WRN-28-10, Learning setup=Self-supervised, External data=false, External labels=false2022.04 | 73.6 | 86.1 | |
| MABASBackbone=ResNet-122021.03 | 73.51 | 85.49 | |
| MetaOptNetBackbone=ResNet-122026.05 | 72.8 | 84.3 | |
| MeaOptNetBackbone=ResNet-122021.03 | 72.6 | 84.3 | |
| MetaOptBackbone=ResNet122022.10 | 72.6 | 84.3 | |
| Prototypical NetworksBackbone=ResNet-122021.03 | 72.2 | 83.5 | |
| ProtoNetBackbone=ResNet-122026.05 | 72.2 | 83.5 | |
| MetaOpNetBackbone=ResNet-122021.06 | 72 | 84.2 | |
| MetaOpt-SVMBackbone=ResNet12, Learning setup=Inductive, External data=false, External labels=false2022.04 | 72 | 84.3 | |
| BECLRBackbone=RN18, Setting=Unsup.2024.02 | 70.39 | 81.56 | |
| Shot-FreeBackbone=ResNet-122021.06 | 69.2 | 84.7 | |
| Shot-FreeBackbone=ResNet122022.10 | 69.2 | 84.7 | |
| R2D2Backbone=Conv-42021.06 | 65.3 | 79.4 | |
| R2D2Backbone=Conv-512F2021.03 | 65.3 | 79.4 | |
| R2D2Backbone=96-192-384-5122022.10 | 65.3 | 79.4 | |
| MAMLBackbone=Conv-42021.06 | 58.9 | 71.5 | |
| MAMLBackbone=32-32-32-322022.10 | 58.9 | 71.5 | |
| ProtoNetBackbone=Conv-42021.06 | 55.5 | 72 | |
| Prototypical NetworksBackbone=Conv-64F2021.03 | 55.5 | 72 | |
| ProtoNetBackbone=64-64-64-642022.10 | 55.5 | 72 | |
| LF2CSBackbone=RN18, Setting=Unsup.2024.02 | 55.04 | 70.62 | |
| RelationNetBackbone=Conv-42021.06 | 55 | 69.3 | |
| Relation NetworksBackbone=Conv-256F2021.03 | 55 | 69.3 | |
| RelaNetBackbone=64-96-128-2562022.10 | 55 | 69.3 | |
| HMSBackbone=RN18, Setting=Unsup.2024.02 | 54.65 | 73.7 | |
| SimCLRBackbone=RN18, Setting=Unsup.2024.02 | 54.56 | 71.19 | |
| MoCo v2Backbone=RN18, Setting=Unsup.2024.02 | 52.73 | 67.81 | |
| BYOLBackbone=RN18, Setting=Unsup.2024.02 | 51.33 | 66.73 | |
| CUMCABackbone=RN18, Setting=Unsup.2024.02 | 50.48 | 67.83 | |
| MoCHiBackbone=RN18, Setting=Unsup.2024.02 | 50.42 | 65.91 | |
| ProtoNetBackbone=CNN-4-64, Learning setup=Inductive, External data=false, External labels=false2022.04 | 49.4 | 68.2 | |
| DeepEMDBackbone=ResNet-122021.06 | 46.47 | 63.22 | |
| FREE5Data recovery process=Inversion, Adaptation steps=5, Time (GPU hours)=0.11h2024.05 | 39.13 | 52.58 | |
| PURERData recovery process=Inversion, Time (GPU hours)=1.21h2024.05 | 38.66 | 51.95 | |
| BiDf-MKDData recovery process=Inversion, Time (GPU hours)=2.47h2024.05 | 37.66 | 51.16 | |
| FREE2Data recovery process=Inversion, Adaptation steps=2, Time (GPU hours)=0.05h2024.05 | 36.56 | 47.31 | |
| DROData recovery process=Non-inversion2024.05 | 30.43 | 36.21 | |
| OTAData recovery process=Non-inversion2024.05 | 29.1 | 34.33 | |
| RandomData recovery process=Non-inversion2024.05 | 28.59 | 34.77 | |
| AverageData recovery process=Non-inversion2024.05 | 23.96 | 27.04 | |
| Best-ModelData recovery process=Non-inversion2024.05 | 21.68 | 25.05 |