Image Classification on ImageNet-C (IN-C and Retention)
32.7IN-C ScoreResNet18
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ResNet18Param.=11M, FLOPs=1.8G2022.04 | 32.7 | 46.8 | — | — | |
| MBV2Param.=4M, FLOPs=0.4G2022.04 | 35 | 47.9 | — | — | |
| PVTV2-B0Param.=3M, FLOPs=0.6G2022.04 | 36.2 | 51.3 | — | — | |
| EffiNet-B0Param.=5M, FLOPs=0.4G2022.04 | 41.1 | 53 | — | — | |
| ResNet50Param.=25M, FLOPs=4.1G2022.04 | 50.6 | 64.1 | — | — | |
| PVTV2-B1Param.=13M, FLOPs=2.1G2022.04 | 51.7 | 65.7 | — | — | |
| Swin-TParam.=28M, FLOPs=4.5G2022.04 | 55.4 | 68.1 | — | — | |
| FAN-T-HybridParam.=7M, FLOPs=3.5G2022.04 | 57.4 | 71.7 | — | — | |
| FAN-T-ViTParam.=7M, FLOPs=1.3G2022.04 | 57.5 | 72.6 | — | — | |
| DeiT-SParam.=22M, FLOPs=4.6G2022.04 | 58.1 | 72.7 | — | — | |
| ConvNeXt-TParam.=29M, FLOPs=4.5G2022.04 | 59.1 | 71.9 | — | — | |
| Swin-SParam.=50M, FLOPs=8.7G2022.04 | 60.4 | 72.8 | — | — | |
| Swin-BParam.=88M, FLOPs=15.4G2022.04 | 60.4 | 72.3 | — | — | |
| ConvNeXt-SParam.=50M, FLOPs=8.7G2022.04 | 61.7 | 74.2 | — | — | |
| ConvNeXt-BParam.=89M, FLOPs=15.4G2022.04 | 61.7 | 73.6 | — | — | |
| DeiT-BParam.=89M, FLOPs=17.6G2022.04 | 62.7 | 76.7 | — | — | |
| FAN-S-ViTParam.=28M, FLOPs=5.3G2022.04 | 64.5 | 77.8 | — | — | |
| FAN-S-HybridParam.=26M, FLOPs=6.7G2022.04 | 64.7 | 77.5 | — | — | |
| FAN-B-HybridParam.=50M, FLOPs=11.3G2022.04 | 66.4 | 79.1 | — | — | |
| FAN-B-ViTParam.=54M, FLOPs=10.4G2022.04 | 67 | 80.1 | — | — | |
| FAN-L-ViTParam.=81M, FLOPs=15.8G2022.04 | 67.7 | 80.7 | — | — | |
| FAN-L-HybridParam.=77M, FLOPs=16.9G2022.04 | 68.3 | 81 | — | — | |
| FAN-B-HybridParam.=50M, FLOPs=11.3G, Pre-trained=ImageNet-22K2022.04 | 70.5 | 82.4 | — | — | |
| FAN-L-HybridParam.=77M, FLOPs=16.9G, Pre-trained=ImageNet-22K2022.04 | 73.6 | 85.1 | — | — | |
| Baseline circuitParadigm=Baseline circuit, Scoring Top-K=Relevance, Backbone=ResNet-1012026.02 | — | — | 74 | 0.7 | |
| Baseline circuitParadigm=Baseline circuit, Scoring Top-K=Activation, Backbone=ResNet-1012026.02 | — | — | 68 | 0.7 | |
| Baseline circuitParadigm=Baseline circuit, Scoring Top-K=Rank, Backbone=ResNet-1012026.02 | — | — | 43 | 0.7 | |
| Certified circuitParadigm=Certified circuit, Scoring Top-K=Relevance, Backbone=ResNet-1012026.02 | — | — | 93 | 0.34 | |
| Certified circuitParadigm=Certified circuit, Scoring Top-K=Activation, Backbone=ResNet-1012026.02 | — | — | 72 | 0.5 | |
| Certified circuitParadigm=Certified circuit, Scoring Top-K=Rank, Backbone=ResNet-1012026.02 | — | — | 71 | 0.58 | |
| ModelParadigm=Unmodified, Scoring Top-K=–, Backbone=ResNet-1012026.02 | — | — | 59 | 1 |