Domain Generalization on PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet
99PACS AccuracySIMPLE+
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SIMPLE+Architecture=ModelPool-B, Pretraining=ModelPool-B, #Param.=> 1,000M, Trainable #Param.=0.9M2023.10 | 99 | 82.7 | 87.7 | 59 | 61.9 | 78.1 | |
| ERM KAdaptation-MoA-EnsembleArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=261.3M2023.10 | 97.6 | 83.4 | 90.9 | 54.3 | 63.1 | 77.9 | |
| ERM KAdaptationArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=85.9M, Trainable #Param.=0.14M2023.10 | 97.5 | 83 | 90.3 | 51.9 | 62.7 | 77.1 | |
| MIRO+Architecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=167.2M, Trainable #Param.=83.6M2023.10 | 97.4 | 79.9 | 80.4 | 58.9 | 53.8 | 74.1 | |
| ERM KAdaptation-MoAArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=87.3M, Trainable #Param.=1.5M2023.10 | 97.4 | 83.1 | 90.6 | 52.8 | 62.7 | 77.3 | |
| ERM LoRA-MoAArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=87.2M, Trainable #Param.=1.5M2023.10 | 96.9 | 82.8 | 89.5 | 49.2 | 62.2 | 75.9 | |
| MIRO+SWAD+Architecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=167.2M, Trainable #Param.=83.6M2023.10 | 96.8 | 81.7 | 83.3 | 64.3 | 60.7 | 77.3 | |
| MIRO*Architecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=172M, Trainable #Param.=85.8M2023.10 | 96.7 | 82.4 | 87.3 | 52.3 | 50.6 | 73.9 | |
| VL2V-SDArchitecture=ViT-B/16, Pretraining=CLIP OpenAI, #Param.=235.8M, Trainable #Param.=83.6M2023.10 | 96.7 | 83.3 | 87.4 | 58.5 | 62.8 | 77.7 | |
| ERM LoRA, r=2Architecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=85.9M, Trainable #Param.=0.11M2023.10 | 96.4 | 82.6 | 86.7 | 46.1 | 61.5 | 74.7 | |
| ERM KAdaptation-MoAArchitecture=ViT-B/16, Pretraining=CLIP OpenAI, #Param.=87.3M, Trainable #Param.=1.5M2023.10 | 96.2 | 83.6 | 84.5 | 54.3 | 59.9 | 75.7 | |
| EoAArchitecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=> 500M2023.10 | 95.8 | 81.1 | 83.9 | 61.1 | 60.9 | 76.6 | |
| MIROArchitecture=ViT-B/16, Pretraining=CLIP OpenAI, #Param.=172M, Trainable #Param.=85.8M2023.10 | 95.6 | 82.2 | 82.5 | 54.3 | 54 | 73.7 | |
| SMAArchitecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=83.6M, Trainable #Param.=83.6M2023.10 | 95.5 | 80.7 | 82 | 59.7 | 60 | 75.6 | |
| VL2V-ADiPArchitecture=ViT-B/16, Pretraining=CLIP OpenAI, #Param.=235.8M, Trainable #Param.=83.6M2023.10 | 94.9 | 81.9 | 85.7 | 55.4 | 59.4 | 75.5 | |
| ERM+SWAD+Architecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=83.6M, Trainable #Param.=83.6M2023.10 | 94.7 | 79.7 | 80 | 57.9 | 53.6 | 73.2 | |
| ERM CompacterArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=85.9M, Trainable #Param.=0.10M2023.10 | 94.1 | 81 | 83 | 35.9 | 56.2 | 70 | |
| ERM AttentionArchitecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=85.8M, Trainable #Param.=28.4M2023.10 | 93.8 | 82 | 85.9 | 51.4 | 57.2 | 74.1 | |
| ERMArchitecture=RegNetY-16GF, Pretraining=SWAG IG3B, #Param.=83.6M, Trainable #Param.=83.6M2023.10 | 89.6 | 78.6 | 71.9 | 51.4 | 48.5 | 68 | |
| CADGPre-training=ImageNet, Ensemble=true2022.03 | 89.1 | 79.6 | 79.9 | 54.2 | 49.8 | 70.5 | |
| EoAPre-training=ImageNet, Ensemble=true2022.03 | 88.6 | 79.1 | 72.5 | 52.3 | 47.4 | 68 | |
| SIMPLEArchitecture=ModelPool-A, Pretraining=ModelPool-A, #Param.=> 1,000M, Trainable #Param.=0.9M2023.10 | 88.6 | 79.9 | 84.6 | 57.6 | 49.2 | 72 | |
| SWADPre-training=ImageNet, Ensemble=true2022.03 | 88.1 | 79.1 | 70.6 | 50 | 46.5 | 66.9 | |
| ERM (Baseline)Architecture=ViT-B/16, Pretraining=CLIP LAION2B, #Param.=85.8M, Trainable #Param.=85.8M2023.10 | 85.8 | 78.5 | 78.1 | 41 | 52.2 | 67.1 | |
| SEDGEPre-training=ImageNet, Ensemble=true2022.03 | 84.1 | 79.8 | 79.9 | 56.8 | 46.3 | 69.4 | |
| ERM+Architecture=ViT-B/16, Pretraining=CLIP OpenAI, #Param.=85.8M, Trainable #Param.=85.8M2023.10 | 83.4 | 75.9 | 66.4 | 35.3 | 44.4 | 61.1 | |
| random ensemblePre-training=ImageNet, Ensemble=true2022.03 | 58.1 | 58.5 | 59.6 | 31.5 | 15.8 | 44.5 |