Image Clustering on STL-10
98.5ACCLAIC framework
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LAIC frameworkBackbone=ViT-B/322026.03 | 98.5 | 96.1 | 96.7 | — | — | |
| KEC_TURTLE2026.04 | 98.5 | 96.1 | 96.7 | — | — | |
| TURTLE (1-space)Training Protocol=1-space2026.04 | 98.4 | 95.8 | 96.4 | — | — | |
| PT+SCAN2021.04 | 98.3 | 95.8 | 96.4 | — | — | |
| GradNorm2025.10 | 98.3 | 95.6 | 96.2 | — | — | |
| GradNormBackbone=ViT-B/322026.03 | 98.3 | 95.6 | 96.2 | — | — | |
| KEC_TAC2026.04 | 98.3 | 95.8 | 96.3 | — | — | |
| TACtrained=true2023.10 | 98.2 | 95.5 | 96.1 | — | — | |
| TACBackbone=ViT-B/322026.03 | 98.2 | 95.5 | 96.1 | — | — | |
| TAC2026.04 | 98.2 | 95.6 | 96.1 | — | — | |
| SIC2023.10 | 98.1 | 95.3 | 95.9 | — | — | |
| SIC2025.10 | 98.1 | 95.3 | 95.9 | — | — | |
| SICBackbone=ViT-B/322026.03 | 98.1 | 95.3 | 95.9 | — | — | |
| MCABackbone=ViT-B/322026.03 | 98.1 | 95.5 | 96 | — | — | |
| PRO-DSCBackbone=ViT-B/322026.03 | 98.1 | 95.4 | 95.9 | — | — | |
| SIC2026.04 | 98.1 | 95.3 | 95.9 | — | — | |
| TAC_TURTLE2026.04 | 98 | 95.3 | 95.7 | — | — | |
| K principal concept identification2021.04 | 97.9 | 95.2 | 95.1 | — | — | |
| KEC (no train)Training Protocol=no train2026.04 | 97.9 | 95.1 | 95.5 | — | — | |
| CLIPEvaluation Protocol=zero-shot2023.10 | 97.1 | 93.9 | 93.7 | — | — | |
| CLIPmode=zero-shot2025.10 | 97.1 | 93.9 | 93.7 | — | — | |
| ZS-CLIPBackbone=ViT-B/32, zero-shot=true2026.03 | 97.1 | 93.9 | 93.7 | — | — | |
| CLIP (zero-shot)Training Protocol=zero-shot2026.04 | 97 | 93.8 | 93.7 | — | — | |
| TemiClustering mode=Offline, Pre-training data=ImageNet-1k2024.07 | 96.7 | 93.6 | 93 | — | — | |
| TAC (no train)Training Protocol=no train2026.04 | 94.8 | 92.5 | 89.9 | — | — | |
| TACtrained=false2023.10 | 94.5 | 92.3 | 89.5 | — | — | |
| TAC2025.10 | 94.5 | 92.3 | 89.5 | — | — | |
| CLIPEvaluation Protocol=k-means2023.10 | 94.3 | 91.7 | 89.1 | — | — | |
| CLIPmode=k-means2025.10 | 94.3 | 91.7 | 89.1 | — | — | |
| CLIP (k-means)Clustering Strategy=k-means2026.04 | 94.3 | 91.8 | 89.2 | — | — | |
| SPICEbackbone=WideResNet-37-2, joint training=true2021.03 | 93.8 | 87.2 | 87 | — | — | |
| SPICESelf-labeling fine-tuning=true, Clustering mode=Online2024.07 | 93.8 | 87.2 | 87 | — | — | |
| DPACOPASelf-labeling fine-tuning=true, Clustering mode=Online2024.07 | 93.4 | 86.3 | 86.1 | — | — | |
| SPICEBackbone=ResNet342021.03 | 92.9 | 0.86 | 0.853 | — | — | |
| ConNREvaluation Protocol=k-means2023.12 | 92.6 | 85.2 | 84.6 | — | — | |
| DPACOPAClustering mode=Online2024.07 | 92.6 | 84.5 | 84.7 | — | — | |
| PT Only2021.04 | 92.1 | 89.4 | 85.5 | — | — | |
| SPICEBackbone=ResNet182021.03 | 92 | 0.852 | 0.836 | — | — | |
| RPSC2025.10 | 92 | 83.8 | 83.4 | — | — | |
| SPICE2022.10 | 90.8 | 81.7 | 81.2 | — | — | |
| SPICEsbackbone=WideResNet-37-2, joint training=false2021.03 | 90.8 | 81.7 | 81.2 | — | — | |
| SPICEClustering mode=Online, Cluster size constraint=true2024.07 | 90.8 | 81.7 | 81.2 | — | — | |
| SPICE2023.10 | 90.8 | 81.7 | 81.2 | — | — | |
| SPICE2025.10 | 90.8 | 81.7 | 81.2 | — | — | |
| DPACpacClustering mode=Offline2024.07 | 89.7 | 81.8 | 80 | — | — | |
| TCL2022.10 | 88.8 | 83.3 | 79.7 | — | — | |
| DLME2022.07 | 88.3 | — | — | — | — | |
| DLME-A1variant=A12022.07 | 87.2 | — | — | — | — | |
| TCLTraining paradigm=directly outputting cluster assignments, Image size=224x2242021.11 | 86.8 | 79.9 | 75.7 | — | — | |
| TCL2023.12 | 86.8 | 79.9 | 75.7 | — | — | |
| TCLSelf-labeling fine-tuning=true, Clustering mode=Online2024.07 | 86.8 | 79.9 | 75.7 | — | — | |
| TCL2025.10 | 86.8 | 79.9 | 75.7 | — | — | |
| TCLBackbone=ResNet2026.03 | 86.8 | 79.9 | 75.7 | — | — | |
| SCAN + RUCSelection Strategy=Confidence2020.12 | 86.7 | — | — | — | — | |
| ProPosTraining paradigm=improving representation learning2021.11 | 86.7 | 75.8 | 73.7 | — | — | |
| RUCSCANBackbone=ResNet182021.03 | 86.7 | — | — | — | — | |
| ProPosEvaluation Protocol=k-means2023.12 | 86.7 | 75.8 | 73.7 | — | — | |
| SCAN + RUCSelection Strategy=Hybrid2020.12 | 86.6 | — | — | — | — | |
| SPICEsBackbone=ResNet182021.03 | 86.2 | 0.756 | 0.732 | — | — | |
| BYOLBackbone=ResNet2026.03 | 86.1 | 75.4 | 71.5 | — | — | |
| LFSSBackbone=ResNet2026.03 | 86.1 | 77.1 | 74 | — | — | |
| DLME-A2variant=A22022.07 | 85.9 | — | — | — | — | |
| SCAN_MoCoBackbone=ResNet182021.03 | 85.5 | 0.758 | 0.721 | — | — | |
| DLME-A3variant=A32022.07 | 85.2 | — | — | — | — | |
| CCStrategy=max-performance2021.05 | 85 | 74.6 | 72.6 | — | — | |
| CC2022.07 | 85 | — | — | — | — | |
| CC2022.10 | 85 | 74.6 | 72.6 | — | — | |
| CCTraining paradigm=directly outputting cluster assignments, Image size=224x2242021.11 | 85 | 76.4 | 72.6 | — | — | |
| CCbackbone=WideResNet-37-22021.03 | 85 | 74.6 | 72.6 | — | — | |
| CC2023.12 | 85 | 76.4 | 72.6 | — | — | |
| CCClustering mode=Online, Cluster size constraint=true2024.07 | 85 | 76.4 | 72.6 | — | — | |
| CC2023.10 | 85 | 76.4 | 72.6 | — | — | |
| CC2025.10 | 85 | 76.4 | 72.6 | — | — | |
| CCBackbone=ResNet2026.03 | 85 | 76.4 | 72.6 | — | — | |
| SCAN + RUCSelection Strategy=Metric2020.12 | 84.7 | — | — | — | — | |
| SeCuSelf-labeling fine-tuning=true, Clustering mode=Online2024.07 | 83.6 | 73.3 | 69.3 | — | — | |
| Inference-Time Attention Engineering (ITAE)Backbone=ViT-B/14 (distilled)2024.10 | 82.76 | 88.18 | 75.94 | — | — | |
| BYOLTraining paradigm=learning general representations2021.11 | 82.5 | 71.3 | 65.7 | — | — | |
| BYOLEvaluation Protocol=k-means2023.12 | 82.5 | 71.3 | 65.7 | — | — | |
| CRLC2022.07 | 81.8 | — | — | — | — | |
| SCAN2022.10 | 81.8 | 70.3 | 66.1 | — | — | |
| CRLC2025.10 | 81.8 | 72.9 | 62.8 | — | — | |
| PCLTraining paradigm=improving representation learning2021.11 | 81.6 | 71.8 | 67 | — | — | |
| SCANexperiment_source=official code2020.12 | 81.4 | — | — | — | — | |
| TCCTraining paradigm=directly outputting cluster assignments2021.11 | 81.4 | 73.2 | 68.9 | — | — | |
| TCC2023.12 | 81.4 | 73.2 | 68.9 | — | — | |
| TCCClustering mode=Online, Cluster size constraint=true2024.07 | 81.4 | 73.2 | 68.9 | — | — | |
| SeCuClustering mode=Online2024.07 | 81.4 | 70.7 | 65.7 | — | — | |
| TCC2023.10 | 81.4 | 73.2 | 68.9 | — | — | |
| TCC2025.10 | 81.4 | 73.2 | 68.9 | — | — | |
| SeCu2025.10 | 81.4 | 70.7 | 65.7 | — | — | |
| TCCBackbone=ResNet2026.03 | 81.4 | 73.2 | 68.9 | — | — | |
| SCAN2021.05 | 80.9 | 69.8 | 64.6 | — | — | |
| SCANTraining paradigm=multi-stage methods2021.11 | 80.9 | 69.8 | 64.6 | — | — | |
| SCANBackbone=ResNet182021.03 | 80.9 | 0.698 | 0.646 | — | — | |
| SCAN2023.12 | 80.9 | 69.8 | 64.6 | — | — | |
| SCANSelf-labeling fine-tuning=true, Clustering mode=Offline2024.07 | 80.9 | 69.8 | 64.6 | — | — | |
| SCAN2023.10 | 80.9 | 69.8 | 64.6 | — | — | |
| SCAN2025.10 | 80.9 | 69.8 | 64.6 | — | — | |
| SCANBackbone=ResNet2026.03 | 80.9 | 69.8 | 64.6 | — | — |