Image Clustering on Imagenet dog-15
90.4NMIMAE-CT
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MAE-CTBackbone=ViT-H/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 90.4 | 94.3 | 87.9 | 90.4 | |
| MAE-CTBackbone=ViT-L/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 87.8 | 84.8 | 81.1 | 87.8 | |
| MAE-CTBackbone=ViT-H/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 86.5 | 83.7 | 79.3 | 86.4 | |
| MAE-CTBackbone=ViT-L/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 85.6 | 84.2 | 76.4 | 85.6 | |
| MSNBackbone=ViT-L/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 84.2 | 83.3 | 75.9 | 84.2 | |
| MSNBackbone=ViT-B/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 83.6 | 81 | 73.7 | 83.6 | |
| DINOBackbone=ViT-B/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 78.9 | 79 | 69.4 | 78.9 | |
| iBOTBackbone=ViT-B/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 78.8 | 77.2 | 68.5 | 78.8 | |
| MAE-CTBackbone=ViT-B/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 77.4 | 78.1 | 65 | 77.3 | |
| iBOTBackbone=ViT-L/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 72.8 | 69.4 | 59.9 | 72.8 | |
| Moco v3Backbone=ViT-B/16, Augmentation strategy=extensive image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 71.9 | 72.7 | 59.3 | 71.8 | |
| MAE-CTBackbone=ViT-B/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 69.1 | 70.8 | 54.3 | 69 | |
| ConCURLStrategy=max-performance2021.05 | 63 | 69.5 | 53.1 | — | |
| SPICEbackbone=ResNet-34, joint training=true2021.03 | 62.7 | 67.5 | 52.6 | — | |
| IDFO2021.05 | 57.9 | 61.2 | 41.4 | — | |
| SPICEsbackbone=ResNet-34, joint training=false2021.03 | 57.2 | 64.6 | 47.9 | — | |
| IDFD2021.05 | 54.6 | 59.1 | 41.3 | — | |
| IDFDStrategy=max-performance2021.05 | 54.6 | 59.1 | 41.3 | — | |
| IDFD2021.03 | 54.6 | 59.1 | 41.3 | — | |
| IDTemperature Setting=tuned2021.05 | 47 | 47.6 | 33.5 | — | |
| IDStrategy=max-performance2021.05 | 47 | 47.6 | 33.5 | — | |
| CCStrategy=max-performance2021.05 | 44.5 | 42.9 | 27.4 | — | |
| CCbackbone=ResNet-342021.03 | 44.5 | 42.9 | 27.4 | — | |
| MiCEStatistic=Best2021.05 | 42.3 | 43.9 | 28.6 | — | |
| MiCEStatistic=Mean2021.05 | 39.4 | 39 | 24.7 | — | |
| PICA2021.05 | 35.2 | 35.2 | 20.1 | — | |
| PICAStrategy=max-performance2021.05 | 35.2 | 35.2 | 20.1 | — | |
| PICA2021.03 | 35.2 | 35.2 | 20.1 | — | |
| MoCoPost-processing=spherical k-means, Statistic=Best2021.05 | 34.7 | 33.8 | 19.7 | — | |
| MoCoPost-processing=spherical k-means, Statistic=Mean2021.05 | 34.2 | 30.8 | 18.4 | — | |
| GATClusterStrategy=max-performance2021.05 | 32.2 | 33.3 | 20 | — | |
| GATCluster2021.03 | 32.2 | 33.3 | 20 | — | |
| DCCM2019.04 | 32.1 | 38.3 | 18.2 | — | |
| DCCM2021.05 | 32.1 | 38.3 | 18.2 | — | |
| DCCM2021.05 | 32.1 | 38.3 | 18.2 | — | |
| DCCMStrategy=max-performance2021.05 | 32.1 | 38.3 | 18.2 | — | |
| DCCM2021.03 | 32.1 | 38.3 | 18.2 | — | |
| AttentionCluster2021.05 | 28.1 | 32.2 | 16.3 | — | |
| IDTemperature Setting=original (0.07)2021.05 | 24.8 | 36.5 | 17.2 | — | |
| DSEC2021.03 | 23.6 | 26.4 | 12.4 | — | |
| DAC2019.04 | 21.9 | 27.5 | 11.1 | — | |
| DAC2021.05 | 21.9 | 27.5 | 11.1 | — | |
| DAC2021.05 | 21.9 | 27.5 | 11.1 | — | |
| DACStrategy=max-performance2021.05 | 21.9 | 27.5 | 11.1 | — | |
| DAC2021.03 | 21.9 | 27.5 | 11.1 | — | |
| MAE (reimpl.)Backbone=ViT-L/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 12.8 | 18.9 | 5.2 | 12.6 | |
| DEC2019.04 | 12.2 | 19.5 | 7.9 | — | |
| DEC2021.05 | 12.2 | 19.5 | 7.9 | — | |
| DEC2021.05 | 12.2 | 19.5 | 7.9 | — | |
| DECStrategy=max-performance2021.05 | 12.2 | 19.5 | 7.9 | — | |
| DEC2021.03 | 12.2 | 19.5 | 7.9 | — | |
| GAN2019.04 | 12.1 | 17.4 | 7.8 | — | |
| GANPost-processing=k-means2021.05 | 12.1 | 17.4 | 7.8 | — | |
| DCGAN2021.03 | 12.1 | 17.4 | 7.8 | — | |
| MAE (reimpl.)Backbone=ViT-H/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 11.7 | 18.7 | 4.7 | 11.6 | |
| VAE2019.04 | 10.7 | 17.9 | 7.9 | — | |
| VAEPost-processing=k-means2021.05 | 10.7 | 17.9 | 7.9 | — | |
| VAE2021.03 | 10.7 | 17.9 | 7.9 | — | |
| AE2019.04 | 10.4 | 18.5 | 7.3 | — | |
| DAE2019.04 | 10.4 | 19 | 7.8 | — | |
| AEPost-processing=k-means2021.05 | 10.4 | 18.5 | 7.3 | — | |
| DAEPost-processing=k-means2021.05 | 10.4 | 19 | 7.8 | — | |
| AE2021.05 | 10.4 | 18.5 | 7.3 | — | |
| AEStrategy=max-performance2021.05 | 10.4 | 18.5 | 7.3 | — | |
| SDAEStrategy=max-performance2021.05 | 10.4 | 19 | 7.8 | — | |
| AE2021.03 | 10.4 | 18.5 | 7.3 | — | |
| SDAE2021.03 | 10.4 | 19 | 7.8 | — | |
| DeCNN2019.04 | 9.8 | 17.5 | 7.3 | — | |
| DeCNNStrategy=max-performance2021.05 | 9.8 | 17.5 | 7.3 | — | |
| DeCNN2021.03 | 9.8 | 17.5 | 7.3 | — | |
| SWWAEPost-processing=k-means2021.05 | 9.4 | 15.9 | 7.6 | — | |
| MAE (reimpl.)Backbone=ViT-B/16, Augmentation strategy=minimal image augmentations, Evaluation protocol=k-means (lowest loss over 100 runs)2023.04 | 8.1 | 16 | 3.2 | 7.9 | |
| K-means2019.04 | 5.5 | 10.5 | 2 | — | |
| k-means2021.05 | 5.5 | 10.5 | 2 | — | |
| k-meansStrategy=max-performance2021.05 | 5.5 | 10.5 | 2 | — | |
| k-means2021.03 | 5.5 | 10.5 | 2 | — | |
| JULE2019.04 | 5.4 | 13.8 | 2.8 | — | |
| JULE2021.05 | 5.4 | 13.8 | 2.8 | — | |
| JULEStrategy=max-performance2021.05 | 5.4 | 13.8 | 2.8 | — | |
| JULE2021.03 | 5.4 | 13.8 | 2.8 | — | |
| NMF2019.04 | 4.4 | 11.8 | 1.6 | — | |
| NMFStrategy=max-performance2021.05 | 4.4 | 11.8 | 1.6 | — | |
| NMF2021.03 | 4.4 | 11.8 | 1.6 | — | |
| SC2019.04 | 3.8 | 11.1 | 1.3 | — | |
| SC2021.05 | 3.8 | 11.1 | 1.3 | — | |
| SCStrategy=max-performance2021.05 | 3.8 | 11.1 | 1.3 | — | |
| SC2021.03 | 3.8 | 11.1 | 1.3 | — | |
| AC2019.04 | 3.7 | 13.9 | 2.1 | — | |
| ACStrategy=max-performance2021.05 | 3.7 | 13.9 | 2.1 | — | |
| AC2021.03 | 3.7 | 13.9 | 2.1 | — |