Deep Metric Learning on SOP (test)
81.4Recall@1Intra-Batch Connections
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Intra-Batch ConnectionsEvaluation Protocol=Standard2021.02 | 81.4 | — | — | — | — | — | |
| PA + MetrixBackbone=ResNet-50, Embedding size=512, Metrix variant=feature mixup2021.06 | 81.3 | — | — | — | 91.7 | 96.9 | |
| ProxyNCA++ + MetrixBackbone=ResNet-50, Embedding size=512, Metrix variant=feature mixup2021.06 | 81.3 | — | — | — | 92.7 | 97.1 | |
| Intra-Batch ConnectionsEvaluation Protocol=Reality Check (Musgrave et al., 2020)2021.02 | 81.1 | — | — | — | — | — | |
| MS + MetrixBackbone=ResNet-50, Embedding size=512, Metrix variant=feature mixup2021.06 | 81 | — | — | — | 92 | 97.2 | |
| ProxyNCA++Backbone=ResNet-50, Embedding size=512, source=reported2021.06 | 80.7 | — | — | — | 92 | 96.7 | |
| RankDim=15362020.03 | 79.8 | 91.3 | 96.3 | 90.4 | — | — | |
| Margin + MSDFBaseline Objective=Margin Loss, beta=1.2, S2SD Variant=Multiscale Self-Distillation with Features, Backbone=ResNet-502020.09 | 79.63 | — | — | 90.7 | — | — | |
| Margin + MSDBaseline Objective=Margin Loss, beta=1.2, S2SD Variant=Multiscale Self-Distillation, Backbone=ResNet-502020.09 | 79.26 | — | — | 90.6 | — | — | |
| Margin + DSDBaseline Objective=Margin Loss, beta=1.2, S2SD Variant=Dual Self-Distillation, Backbone=ResNet-502020.09 | 79.05 | — | — | 90.52 | — | — | |
| Margin + MSDFABaseline Objective=Margin Loss, beta=1.2, S2SD Variant=Multiscale Self-Distillation with Features and Affinities, Backbone=ResNet-502020.09 | 78.82 | — | — | 90.49 | — | — | |
| Multisimilarity + MSDFBaseline Objective=Multisimilarity Loss, S2SD Variant=Multiscale Self-Distillation with Features, Backbone=ResNet-502020.09 | 78.59 | — | — | 90.09 | — | — | |
| R-Margin + MSDFBaseline Objective=Relaxed Margin Loss, beta=0.6, S2SD Variant=Multiscale Self-Distillation with Features, Backbone=ResNet-502020.09 | 78.57 | — | — | 90.58 | — | — | |
| MarginBaseline Objective=Margin Loss, beta=1.2, Backbone=ResNet-502020.09 | 78.43 | — | — | 90.4 | — | — | |
| Multisimilarity + MSDBaseline Objective=Multisimilarity Loss, S2SD Variant=Multiscale Self-Distillation, Backbone=ResNet-502020.09 | 78.42 | — | — | 90.09 | — | — | |
| SOFT-TRIPLEDim=5122020.03 | 78.3 | 90.3 | 95.9 | 92 | — | — | |
| Multisimilarity + DSDBaseline Objective=Multisimilarity Loss, S2SD Variant=Dual Self-Distillation, Backbone=ResNet-502020.09 | 78.23 | — | — | 90.08 | — | — | |
| Multisimilarity + MSDFABaseline Objective=Multisimilarity Loss, S2SD Variant=Multiscale Self-Distillation with Features and Affinities, Backbone=ResNet-502020.09 | 78.07 | — | — | 89.88 | — | — | |
| R-Margin + MSDBaseline Objective=Relaxed Margin Loss, beta=0.6, S2SD Variant=Multiscale Self-Distillation, Backbone=ResNet-502020.09 | 78 | — | — | 90.47 | — | — | |
| R-Margin + MSDFABaseline Objective=Relaxed Margin Loss, beta=0.6, S2SD Variant=Multiscale Self-Distillation with Features and Affinities, Backbone=ResNet-502020.09 | 78 | — | — | 90.41 | — | — | |
| MultisimilarityBaseline Objective=Multisimilarity Loss, Backbone=ResNet-502020.09 | 77.99 | — | — | 90 | — | — | |
| R-Margin + DSDBaseline Objective=Relaxed Margin Loss, beta=0.6, S2SD Variant=Dual Self-Distillation, Backbone=ResNet-502020.09 | 77.86 | — | — | 90.5 | — | — | |
| R-MarginBaseline Objective=Relaxed Margin Loss, beta=0.6, Backbone=ResNet-502020.09 | 77.58 | — | — | 90.42 | — | — | |
| MICDim=1282020.03 | 77.2 | 89.4 | 95.6 | 90 | — | — | |
| Ours (Margin + PADS)Dim=1282020.03 | 76.5 | 89 | 95.4 | 89.9 | — | — | |
| ABEDim=5122020.03 | 76.3 | 88.4 | 94.8 | — | — | — | |
| D&CDim=1282020.03 | 75.9 | 88.4 | 94.9 | 90.2 | — | — | |
| HTLDim=5122020.03 | 74.8 | 88.3 | 94.8 | — | — | — | |
| A-BIERDim=5122020.03 | 74.2 | 86.9 | 94 | — | — | — | |
| MarginDim=1282020.03 | 72.7 | 86.2 | 93.8 | 90.8 | — | — | |
| DVMLDim=5122020.03 | 70.2 | 85.2 | 93.8 | 90.8 | — | — | |
| HDMLDim=5122020.03 | 68.7 | 83.2 | 92.4 | 89.3 | — | — |