Image Retrieval on Paris Revisited (Medium)
89.5mAPResNet101-R-MAC+DFS
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ResNet101-R-MAC+DFSBackbone=ResNet-101, Pooling=R-MAC, Re-ranking=DFS (Global Diffusion)2017.11 | 89.5 | 96.7 | |
| ResNet101-GeM+DFSBackbone=ResNet-101, Pooling=GeM, Re-ranking=DFS (Global Diffusion)2017.11 | 88.9 | 96.9 | |
| CiDeR-FTTraining Set=RGLDv2-clean, Fine-tuned=true2024.04 | 87.4 | 97 | |
| CiDeRTraining Set=RGLDv2-clean, Fine-tuned=false2024.04 | 84.6 | 96.7 | |
| CiDeR-FTTraining Set=SfM-120k, Fine-tuned=true2024.04 | 84.5 | 98 | |
| Yokoo et al.Training Set=RGLDv2-clean, Fine-tuned=false2024.04 | 84.1 | 95.4 | |
| ResNet50-DELF–HQE+SVBackbone=ResNet-50, Local Feature=DELF, Query Expansion=HQE, Spatial Verification=true2017.11 | 84 | 98.3 | |
| DOLGTraining Set=RGLDv2-clean, Fine-tuned=false2024.04 | 83.2 | 95.4 | |
| GLAMTraining Set=RGLDv2-clean, Fine-tuned=false2024.04 | 83 | 95 | |
| HOW N/DBackbone=ResNet-18, Paired day-night training data=true2023.09 | 81.4 | — | |
| ResNet101-GeM+αQEBackbone=ResNet-101, Pooling=GeM, Query Expansion=αQE2017.11 | 80.7 | 98.9 | |
| SOLARTraining Set=RGLDv2-clean, Fine-tuned=true2024.04 | 80.3 | 94.6 | |
| CycleGAN CDBackbone=ResNet-18, Paired day-night training data=false2023.09 | 79.8 | — | |
| HED^N GAN CDBackbone=ResNet-18, Paired day-night training data=false2023.09 | 79.7 | — | |
| CiDeRTraining Set=SfM-120k, Fine-tuned=false2024.04 | 79.4 | 97.9 | |
| HOWBackbone=ResNet-18, Paired day-night training data=false2023.09 | 79.4 | — | |
| AGEMTraining Set=SfM-120k, Fine-tuned=true2024.04 | 78.1 | — | |
| GLAMTraining Set=SfM-120k, Fine-tuned=false2024.04 | 77.5 | — | |
| Filip et al.Training Set=SfM-120k, Fine-tuned=true2024.04 | 77.2 | 98.1 | |
| GeMBackbone=ResNet-101, Paired day-night training data=false2023.09 | 76.7 | — | |
| HED^N GAN CDBackbone=ResNet-101, Paired day-night training data=false2023.09 | 76.6 | — | |
| CycleGAN CDBackbone=ResNet-101, Paired day-night training data=false2023.09 | 76.4 | — | |
| TokenTraining Set=RGLDv2-clean, Fine-tuned=false2024.04 | 75.8 | 94.3 | |
| DINOArch.=ViT-S/16, Pretrain Dataset=GLDv22021.04 | 75.3 | — | |
| CIConvBackbone=ResNet-101, Paired day-night training data=false2023.09 | 74.7 | — | |
| RN101+R-MACPretrain=Supervised, Pretrain Dataset=ImNet2021.04 | 74 | — | |
| HesAffNet-HardNet++-HQE+SVDetector=HesAffNet (Hessian + AffNet), Descriptor=HardNet++, Query Expansion=HQE, Spatial Verification=true2017.11 | 73.1 | 98.1 | |
| HesAffNet-HardNet++-HQEDetector=HesAffNet (Hessian + AffNet), Descriptor=HardNet++, Query Expansion=HQE2017.11 | 72.6 | 98.1 | |
| CycleGAN CDBackbone=VGG-16, Paired day-night training data=false2023.09 | 71 | — | |
| SOLARTraining Set=SfM-120k, Fine-tuned=true2024.04 | 70.9 | 98.1 | |
| HED^N GAN CDBackbone=VGG-16, Paired day-night training data=false2023.09 | 70.3 | — | |
| HesAff-RootSIFT–HQE+SVDetector=Hessian-Affine, Descriptor=RootSIFT, Query Expansion=HQE, Spatial Verification=true2017.11 | 70.2 | 98.6 | |
| CLAHETraining Dataset=Retrieval-SfM-N/D, Descriptor Dimension=512D2019.08 | 70 | — | |
| CLAHE N/D CBackbone=VGG-16, Paired day-night training data=true2023.09 | 70 | — | |
| U-Net normalizationTraining Dataset=Retrieval-SfM-N/D, Fine-tuning Setup=normalization, Descriptor Dimension=512D2019.08 | 69.9 | — | |
| VGG GeMTraining Dataset=Retrieval-SfM-N/D, Descriptor Dimension=512D2019.08 | 69.8 | — | |
| CLAHETraining Dataset=Retrieval-SfM, Descriptor Dimension=512D2019.08 | 69.8 | — | |
| GeM N/DBackbone=VGG-16, Paired day-night training data=true2023.09 | 69.8 | — | |
| CLAHE CBackbone=VGG-16, Paired day-night training data=false2023.09 | 69.8 | — | |
| U-Net normalizationTraining Dataset=Retrieval-SfM, Fine-tuning Setup=joint, Descriptor Dimension=512D2019.08 | 69.7 | — | |
| U-Net normalizationTraining Dataset=Retrieval-SfM-N/D, Fine-tuning Setup=joint, Descriptor Dimension=512D2019.08 | 69.6 | — | |
| VGG GeMTraining Dataset=Retrieval-SfM, Descriptor Dimension=512D2019.08 | 69.3 | — | |
| GeMBackbone=VGG-16, Paired day-night training data=false2023.09 | 69.3 | — | |
| HesAff-RootSIFT-HQEDetector=Hessian-Affine, Descriptor=RootSIFT, Query Expansion=HQE2017.11 | 68.9 | 97.3 | |
| EdgeMAC+VGGTraining Dataset=Retrieval-SfM, Descriptor Dimension=1024D2019.08 | 68.8 | — | |
| Gamma correctionTraining Dataset=Retrieval-SfM-N/D, Descriptor Dimension=512D2019.08 | 68.7 | — | |
| Histogram equalizationTraining Dataset=Retrieval-SfM-N/D, Descriptor Dimension=512D2019.08 | 68.3 | — | |
| U-Net normalizationTraining Dataset=Retrieval-SfM-N/D, Fine-tuning Setup=embedding, Descriptor Dimension=512D2019.08 | 68.3 | — | |
| DIRTraining Set=SfM-120k, Fine-tuned=false2024.04 | 68.3 | 97.7 | |
| DINO-vMFBackbone=ViT-Base/162024.05 | 66.7 | — | |
| iBOT-vMFBackbone=ViT-Base/162024.05 | 65.4 | — | |
| GeMTraining Set=SfM-120k, Fine-tuned=false2024.04 | 64.3 | 92.6 | |
| iBOTBackbone=ViT-Base/162024.05 | 64.1 | — | |
| DINOBackbone=ViT-Base/162024.05 | 63.7 | — | |
| DINOArch.=ViT-S/16, Pretrain Dataset=ImNet2021.04 | 63.1 | — | |
| DINOBackbone=ViT-Small/162024.05 | 63.1 | — | |
| ViT-S/16Pretrain=Supervised, Pretrain Dataset=ImNet2021.04 | 63 | — | |
| DINO-vMFBackbone=ViT-Small/162024.05 | 62.1 | — | |
| MSNBackbone=ViT-Small/162024.05 | 61.5 | — | |
| DOLGTraining Set=SfM-120k, Fine-tuned=false2024.04 | 56.6 | 91.1 | |
| MSNBackbone=ViT-Base/162024.05 | 56.3 | — | |
| DINOArch.=ResNet-50, Pretrain Dataset=ImNet2021.04 | 55.9 | — | |
| EdgeMACTraining Dataset=Retrieval-SfM, Descriptor Dimension=512D2019.08 | 43.5 | — |