Image Retrieval on Holidays
94.2mAPDeep Image Retrieval (ResNet50)
Evaluation Results
| Method | Links | |
|---|---|---|
| Deep Image Retrieval (ResNet50)Dim.=2048, Retrieval Protocol=Global descriptors, Backbone=ResNet50, Manual Rotation=true2016.04 | 94.2 | |
| XCIT-S12/8Aggregation strategy=ASMK, k=655362021.06 | 92.3 | |
| XCIT-S12/16Aggregation strategy=ASMK, k=655362021.06 | 92 | |
| Deep Image Retrieval (ResNet50)Dim.=2048, Retrieval Protocol=Global descriptors, Backbone=ResNet502016.04 | 90.7 | |
| XCiT-S12/16Aggregation strategy=VLAD, k=10242021.06 | 90.7 | |
| XCIT-S12/8Aggregation strategy=ASMK, k=10242021.06 | 90.4 | |
| XCiT-S12/16Aggregation strategy=ASMK, k=10242021.06 | 90.4 | |
| Azizpour et al.Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 90 | |
| XCIT-S12/16Aggregation strategy=VLAD, k=2562021.06 | 89.9 | |
| Xinchao et al.Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 89.2 | |
| Deep Image Retrieval (VGG16)Dim.=512, Retrieval Protocol=Global descriptors, Backbone=VGG16, Manual Rotation=true2016.04 | 89.1 | |
| XCIT-S12/8Aggregation strategy=class token, Training protocol=self-supervised DINO2021.06 | 88.3 | |
| Resnet50Aggregation strategy=ASMK, k=655362021.06 | 87.9 | |
| Multigrain-resnet50Aggregation strategy=ASMK, k=10242021.06 | 87.9 | |
| Arandjelovic et al.Dim.=4096, Retrieval Protocol=Global descriptors, Manual Rotation=true2016.04 | 87.5 | |
| XCIT-S12/16Aggregation strategy=class token, Training protocol=self-supervised DINO2021.06 | 87.3 | |
| Tolias et al.Dim.=512, Retrieval Protocol=Global descriptors, Reimplementation=true, Manual Rotation=true2016.04 | 86.9 | |
| Deep Image Retrieval (VGG16)Dim.=512, Retrieval Protocol=Global descriptors, Backbone=VGG162016.04 | 86.7 | |
| XCiT-S12/8Aggregation strategy=class token, Resolution=768 (default)2021.06 | 86.4 | |
| XCIT-S12/16Aggregation strategy=class token, Resolution=10242021.06 | 86.3 | |
| XCIT-S12/16Aggregation strategy=class token, Resolution=768 (default)2021.06 | 86 | |
| XCIT-S12/16Aggregation strategy=class token, Resolution=7682021.06 | 86 | |
| Resnet50Aggregation strategy=ASMK, k=10242021.06 | 86 | |
| Tolias et al.Dim.=512, Retrieval Protocol=Global descriptors, Reimplementation=true2016.04 | 85.2 | |
| Kalantidis et al.Dim.=512, Retrieval Protocol=Global descriptors2016.04 | 84.9 | |
| Perronnin & LarlusDim.=4000, Retrieval Protocol=Global descriptors2016.04 | 84.7 | |
| Deng et al.Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 84.7 | |
| XCIT-S12/16Aggregation strategy=class token, Resolution=5122021.06 | 84.6 | |
| Ng et al.Dim.=128, Retrieval Protocol=Global descriptors2016.04 | 83.6 | |
| XCiT-S12/16Aggregation strategy=class token, Resolution=3842021.06 | 83.4 | |
| Arandjelovic et al.Dim.=4096, Retrieval Protocol=Global descriptors2016.04 | 83.1 | |
| Radenovic et al.Dim.=512, Retrieval Protocol=Global descriptors, Manual Rotation=true2016.04 | 82.5 | |
| ImageNet-21K PretrainingPre-training=ImageNet-21K2021.04 | 82.1 | |
| NetVLADMode=off-the-shelf, Whitening=learnt on Pittsburgh, Representation Dimension=256-D2015.11 | 82.1 | |
| Tolias et al. (QE)Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 81.3 | |
| ImageNet-1K PretrainingPre-training=ImageNet-1K2021.04 | 81.1 | |
| Gong et al.Dim.=2048, Retrieval Protocol=Global descriptors2016.04 | 80.8 | |
| Sum pooling (SPoC w/o center prior)Descriptor Dimension=2562015.10 | 80.2 | |
| Babenko & LempitskyDim.=256, Retrieval Protocol=Global descriptors, Manual Rotation=true2016.04 | 80.2 | |
| NetVLADMode=trained, Representation Dimension=256-D2015.11 | 79.9 | |
| Paulin et al.Dim.=256K, Retrieval Protocol=Global descriptors2016.04 | 79.3 | |
| Gordo et al.Dim.=512, Retrieval Protocol=Global descriptors2016.04 | 79 | |
| Deep fully connected + fine-tuningDescriptor Dimension=2562015.10 | 78.9 | |
| Babenko et al.Dim.=128, Retrieval Protocol=Global descriptors, Manual Rotation=true2016.04 | 78.9 | |
| Tao et al.Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 78.7 | |
| Deep fully connected + VLADDescriptor Dimension=5122015.10 | 78.3 | |
| Gordo et al.Representation Dimension=256-D2015.11 | 78.3 | |
| Shen et al.Retrieval Protocol=Matching / Spatial verif. / QE2016.04 | 76.2 | |
| Babenko et al.Dim.=128, Retrieval Protocol=Global descriptors2016.04 | 75.9 | |
| Mikulik et al.Retrieval Protocol=Matching / Spatial verif. / QE, Manual Rotation=true2016.04 | 75.8 | |
| Deep fully connectedDescriptor Dimension=2562015.10 | 74.9 | |
| Razavian et al.Representation Dimension=256-D2015.11 | 74.2 | |
| SIFT + Triang. + Democr. aggr.Descriptor Dimension=10242015.10 | 72 | |
| Jégou & ZissermanDim.=1024, Retrieval Protocol=Global descriptors2016.04 | 72 | |
| Deep convolutional + Max poolingDescriptor Dimension=2562015.10 | 71.6 | |
| XCIT-S12/16Aggregation strategy=class token, Resolution=2242021.06 | 71.5 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=best2014.12 | 70.2 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=5122014.12 | 69.8 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=20482014.12 | 68.6 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=1282014.12 | 68.5 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=best2014.12 | 66.5 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=5122014.12 | 66.3 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=best2014.12 | 66 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=5122014.12 | 65.8 | |
| Jégou and ZissermanRepresentation Dimension=256-D2015.11 | 65.7 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=1282014.12 | 65.3 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=20482014.12 | 65 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=1282014.12 | 64.7 | |
| VLAD+adapt+innormK=256, D=16384, Dimension reduction (D')=best2014.12 | 64.6 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=20482014.12 | 63.7 | |
| FVK=256, D=16384, Dimension reduction (D')=best2014.12 | 62.6 | |
| FVK=256, D=16384, Dimension reduction (D')=20482014.12 | 62.6 | |
| VLAD+adapt+innormK=256, D=16384, Dimension reduction (D')=1282014.12 | 62.5 | |
| K-meansNetK=256, D=8192, Dimension reduction (D')=best2014.12 | 62.5 | |
| K-meansNetK=256, D=8192, Dimension reduction (D')=5122014.12 | 62.5 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=642014.12 | 62.5 | |
| VLADK=256, D=16384, Dimension reduction (D')=best2014.12 | 62.1 | |
| VLADK=256, D=16384, Dimension reduction (D')=20482014.12 | 62.1 | |
| SIFT + Triang. + Democr. aggr.Descriptor Dimension=1282015.10 | 61.7 | |
| Jégou & ZissermanDim.=128, Retrieval Protocol=Global descriptors2016.04 | 61.7 | |
| K-meansNetK=256, D=8192, Dimension reduction (D')=1282014.12 | 61.3 | |
| K-meansNetK=256, D=8192, Dimension reduction (D')=20482014.12 | 59.8 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=642014.12 | 59.3 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=642014.12 | 58.3 | |
| C-SVDDK=3200, D=12800, Dimension reduction (D')=best2014.12 | 57.4 | |
| FVK=256, D=16384, Dimension reduction (D')=5122014.12 | 57 | |
| C-SVDDK=3200, D=12800, Dimension reduction (D')=5122014.12 | 57 | |
| C-SVDDK=3200, D=12800, Dimension reduction (D')=20482014.12 | 56.8 | |
| VLADK=256, D=16384, Dimension reduction (D')=5122014.12 | 56.7 | |
| K-meansNetK=256, D=8192, Dimension reduction (D')=642014.12 | 55.5 | |
| K-meansK=3200, D=12800, Dimension reduction (D')=best2014.12 | 55.2 | |
| K-meansK=3200, D=12800, Dimension reduction (D')=5122014.12 | 54.9 | |
| K-meansK=3200, D=12800, Dimension reduction (D')=20482014.12 | 54.5 | |
| VLADK=256, D=16384, Dimension reduction (D')=1282014.12 | 54.2 | |
| FVK=256, D=16384, Dimension reduction (D')=1282014.12 | 53.8 | |
| MSRV+C-SVDDNetK=256, D=8192×8, Dimension reduction (D')=322014.12 | 53.8 | |
| C-SVDDK=3200, D=12800, Dimension reduction (D')=1282014.12 | 53.1 | |
| C-SVDDNetK=256, D=8192, Dimension reduction (D')=322014.12 | 52.1 | |
| K-meansK=3200, D=12800, Dimension reduction (D')=1282014.12 | 51.6 | |
| MSRV+K-meansNetK=256, D=8192×8, Dimension reduction (D')=322014.12 | 51.5 |