Image Retrieval on Cars-196
97Recall@1SAGA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SAGAProtocol=Zero-shot, Backbone=Qwen3-VL-8B, Resolution=224x224, Pooling Method=learned attention pooler2026.06 | 97 | — | 98.6 | 89 | |
| Potential FieldBackbone=DINO, Embedding Dimension=3842024.05 | 94.7 | 96.5 | 97.8 | — | |
| Potential FieldBackbone=DINO, Embedding Dimension=384, Variant=PF, Training Setting=Original Paper Reported (†)2026.01 | 94.7 | 96.5 | 97.8 | — | |
| PFProtocol=Zero-shot, Backbone=Qwen3-VL-8B, Resolution=224x224, Pooling Method=mean pooling2026.06 | 93.7 | — | 97.8 | 86 | |
| PAProtocol=Zero-shot, Backbone=Qwen3-VL-8B, Resolution=224x224, Pooling Method=mean pooling2026.06 | 93.4 | — | 97.3 | 84 | |
| BLenDeR - PABackbone=DINO, Embedding Dimension=384, Base Method=Proxy Anchor2026.01 | 93 | 96.2 | 97.8 | — | |
| DiVABackbone=ResNet50, Embedding Dimension=5122024.05 | 92.9 | — | — | — | |
| Potential FieldBackbone=ResNet50, Embedding Dimension=5122024.05 | 92.7 | 95.5 | 97.6 | — | |
| Potential FieldBackbone=ResNet50, Embedding Dimension=512, Variant=PF, Training Setting=Original Paper Reported (†)2026.01 | 92.7 | 95.5 | 97.6 | — | |
| BLenDeR - PABackbone=ResNet50, Embedding Dimension=512, Base Method=Proxy Anchor2026.01 | 92.3 | 95.6 | 97.3 | — | |
| Proxy AnchorBackbone=DINO, Embedding Dimension=384, Variant=PA, Training Setting=Replicated (*)2026.01 | 92.1 | 95.6 | 97.6 | — | |
| PA+niVBackbone=BN Inception, Embedding Dimension=5122024.05 | 92 | — | — | — | |
| BLenDeR - PFBackbone=ResNet50, Embedding Dimension=512, Base Method=Potential Field2026.01 | 91.9 | 95.6 | 97.7 | — | |
| Potential FieldBackbone=ViT, Embedding Dimension=3842024.05 | 91.5 | 95.2 | 97.4 | — | |
| BLenDeR - PFBackbone=DINO, Embedding Dimension=384, Base Method=Potential Field2026.01 | 91.5 | 95.6 | 97.6 | — | |
| Potential FieldBackbone=ViT, Embedding Dimension=384, Variant=PF, Training Setting=Original Paper Reported (†)2026.01 | 91.5 | 95.2 | 97.4 | — | |
| HIERBackbone=DINO, Embedding Dimension=3842024.05 | 91.3 | 95.2 | 97.1 | — | |
| HIERBackbone=DINO, Embedding Dimension=3842026.01 | 91.3 | 95.2 | 97.1 | — | |
| Potential FieldBackbone=DINO, Embedding Dimension=384, Variant=PF, Training Setting=Replicated (*)2026.01 | 90.8 | 95 | 97.3 | — | |
| DiVABackbone=BN Inception, Embedding Dimension=5122024.05 | 90.7 | — | — | — | |
| Proxy AnchorBackbone=ResNet50, Embedding Dimension=512, Variant=PA, Training Setting=Replicated (*)2026.01 | 90.5 | 94.5 | 96.8 | — | |
| Potential FieldBackbone=ResNet50, Embedding Dimension=512, Variant=PF, Training Setting=Replicated (*)2026.01 | 90.2 | 94.3 | 96.9 | — | |
| Potential FieldBackbone=BN Inception, Embedding Dimension=5122024.05 | 90.1 | 93.9 | 96.3 | — | |
| HISTBackbone=ResNet50, Embedding Dimension=5122024.05 | 89.6 | 93.9 | 96.4 | — | |
| HSE-PABackbone=ResNet50, Embedding Dimension=5122024.05 | 89.6 | 93.8 | 96 | — | |
| HISTBackbone=ResNet50, Embedding Dimension=5122026.01 | 89.6 | 93.9 | 96.4 | — | |
| HSE-PABackbone=ResNet50, Embedding Dimension=5122026.01 | 89.6 | 93.8 | 96 | — | |
| HypBackbone=DINO, Embedding Dimension=3842024.05 | 89.2 | 94.1 | 96.7 | — | |
| HypBackbone=DINO, Embedding Dimension=3842026.01 | 89.2 | 94.1 | 96.7 | — | |
| DFML-PABackbone=BN Inception, Embedding Dimension=5122024.05 | 88.4 | — | — | — | |
| BLenDeR - PABackbone=ViT, Embedding Dimension=384, Base Method=Proxy Anchor2026.01 | 88.4 | 93.6 | 96.4 | — | |
| HIERBackbone=ViT, Embedding Dimension=3842024.05 | 88.3 | 93.2 | 96.1 | — | |
| HIERBackbone=ViT, Embedding Dimension=3842026.01 | 88.3 | 93.2 | 96.1 | — | |
| HIERBackbone=ResNet50, Embedding Dimension=5122024.05 | 88.2 | 93 | 95.6 | — | |
| HIERBackbone=ResNet50, Embedding Dimension=5122026.01 | 88.2 | 93 | 95.6 | — | |
| MS+DASBackbone=ResNet50, Embedding Dimension=5122024.05 | 87.8 | 93.1 | 96 | — | |
| MS+DASBackbone=ResNet50, Embedding Dimension=5122026.01 | 87.8 | 93.1 | 96 | — | |
| Proxy AnchorBackbone=ResNet50, Embedding Dimension=5122024.05 | 87.7 | 92.9 | 95.8 | — | |
| Proxy AnchorBackbone=ResNet50, Embedding Dimension=512, Variant=PA2026.01 | 87.7 | 92.9 | 95.8 | — | |
| BLenDeR - PFBackbone=ViT, Embedding Dimension=384, Base Method=Potential Field2026.01 | 87.7 | 93.3 | 96.4 | — | |
| DiVABackbone=ResNet50, Embedding Dimension=5122026.01 | 87.6 | 92.9 | — | — | |
| HISTBackbone=BN Inception, Embedding Dimension=5122024.05 | 87.4 | 92.5 | 95.4 | — | |
| Proxy AnchorBackbone=ViT, Embedding Dimension=384, Variant=PA, Training Setting=Replicated (*)2026.01 | 87.2 | 92.6 | 96 | — | |
| DRML-PABackbone=BN Inception, Embedding Dimension=5122024.05 | 86.9 | 92.1 | 95.2 | — | |
| HypBackbone=ViT, Embedding Dimension=3842024.05 | 86.5 | 92.1 | 95.3 | — | |
| HypBackbone=ViT, Embedding Dimension=3842026.01 | 86.5 | 92.1 | 95.3 | — | |
| Potential FieldBackbone=ViT, Embedding Dimension=384, Variant=PF, Training Setting=Replicated (*)2026.01 | 86.3 | 92.4 | 96.1 | — | |
| Proxy AnchorBackbone=BN Inception, Embedding Dimension=5122024.05 | 86.1 | 91.7 | 95 | — | |
| ARCBackbone=SigLip 2 (ViT-B/16), Compression Ratio=50%2026.02 | 85.9 | — | — | — | |
| MS+DASBackbone=BN Inception, Embedding Dimension=5122024.05 | 85.7 | 91.6 | 95.3 | — | |
| BaselineBackbone=SigLip 2 (ViT-B/16), Compression Ratio=0%2026.02 | 85.7 | — | — | — | |
| ProxyGMLBackbone=BN Inception, Embedding Dimension=5122024.05 | 85.5 | 91.8 | 95.3 | — | |
| AutoEncoder-VAEBackbone=SigLip 2 (ViT-B/16), Compression Ratio=50%2026.02 | 85.5 | — | — | — | |
| ARCBackbone=SigLip 2 (ViT-B/16), Compression Ratio=75%2026.02 | 85.5 | — | — | — | |
| DCML-MDWBackbone=ResNet50, Embedding Dimension=5122024.05 | 85.2 | 91.8 | 96 | — | |
| CPMLBackbone=ResNet50, Embedding Dimension=5122024.05 | 85.2 | 91.5 | 95.2 | — | |
| Q-FormerBackbone=SigLip 2 (ViT-B/16), Compression Ratio=50%2026.02 | 85.2 | — | — | — | |
| DCML-MDWBackbone=ResNet50, Embedding Dimension=5122026.01 | 85.2 | 91.8 | 96 | — | |
| Proxy NCA++Backbone=ResNet50, Embedding Dimension=5122024.05 | 85.1 | — | — | — | |
| PTQ-P4QBackbone=SigLip 2 (ViT-B/16), Compression Ratio=75%, Bits=8-bit, Fine-tuning=true2026.02 | 85.1 | — | — | — | |
| AutoEncoder-VAEBackbone=SigLip 2 (ViT-B/16), Compression Ratio=75%2026.02 | 84.9 | — | — | — | |
| PTQ-MinMaxBackbone=SigLip 2 (ViT-B/16), Compression Ratio=75%, Bits=8-bit, Fine-tuning=true2026.02 | 84.8 | — | — | — | |
| ARCBackbone=SigLip 2 (ViT-B/16), Compression Ratio=87.5%2026.02 | 84.6 | — | — | — | |
| SoftTripleBackbone=BN Inception, Embedding Dimension=5122024.05 | 84.5 | 90.7 | 94.5 | — | |
| Q-FormerBackbone=SigLip 2 (ViT-B/16), Compression Ratio=75%2026.02 | 84.3 | — | — | — | |
| N.SoftmaxBackbone=ResNet50, Embedding Dimension=5122024.05 | 84.2 | 90.4 | 94.4 | — | |
| N.SoftmaxBackbone=ResNet50, Embedding Dimension=5122026.01 | 84.2 | 90.4 | 94.4 | — | |
| MultiSimilarityBackbone=BN Inception, Embedding Dimension=5122024.05 | 84.1 | 90.4 | 94 | — | |
| CircleLossBackbone=BN Inception, Embedding Dimension=5122024.05 | 83.4 | 89.8 | 94.1 | — | |
| ESPHNBackbone=ResNet50, Embedding Dimension=5122024.05 | 82.7 | 89.3 | 93 | — | |
| AutoEncoder-VAEBackbone=SigLip 2 (ViT-B/16), Compression Ratio=87.5%2026.02 | 82.7 | — | — | — | |
| ESPHNBackbone=ResNet50, Embedding Dimension=5122026.01 | 82.7 | 89.3 | 93 | — | |
| HSE-MBackbone=BN Inception, Embedding Dimension=5122024.05 | 82 | 88.9 | 93.3 | — | |
| ARCBackbone=SigLip 2 (ViT-B/16), Compression Ratio=93.75%2026.02 | 81.8 | — | — | — | |
| PTQ-P4QBackbone=SigLip 2 (ViT-B/16), Compression Ratio=87.5%, Bits=4-bit, Fine-tuning=true2026.02 | 81.7 | — | — | — | |
| Q-FormerBackbone=SigLip 2 (ViT-B/16), Compression Ratio=87.5%2026.02 | 81.6 | — | — | — | |
| HTLBackbone=BN Inception, Embedding Dimension=5122024.05 | 81.4 | 88 | 92.7 | — | |
| PTQ-MinMaxBackbone=SigLip 2 (ViT-B/16), Compression Ratio=87.5%, Bits=4-bit, Fine-tuning=true2026.02 | 80.9 | — | — | — | |
| AutoEncoder-VAEBackbone=SigLip 2 (ViT-B/16), Compression Ratio=93.75%2026.02 | 80.5 | — | — | — | |
| ARCBackbone=CN-CLIP (ViT-H/14), Compression Ratio=50%2026.02 | 80.4 | — | — | — | |
| Q-FormerBackbone=SigLip 2 (ViT-B/16), Compression Ratio=93.75%2026.02 | 79.9 | — | — | — | |
| BaselineBackbone=CN-CLIP (ViT-H/14), Compression Ratio=0%2026.02 | 79.8 | — | — | — | |
| ARCBackbone=CN-CLIP (ViT-H/14), Compression Ratio=75%2026.02 | 79.8 | — | — | — | |
| PTQ-P4QBackbone=CN-CLIP (ViT-H/14), Compression Ratio=75%, Bits=8-bit, Fine-tuning=true2026.02 | 79.7 | — | — | — | |
| PTQ-MinMaxBackbone=CN-CLIP (ViT-H/14), Compression Ratio=75%, Bits=8-bit, Fine-tuning=true2026.02 | 79.5 | — | — | — | |
| AutoEncoder-VAEBackbone=CN-CLIP (ViT-H/14), Compression Ratio=50%2026.02 | 79.4 | — | — | — | |
| Q-FormerBackbone=CN-CLIP (ViT-H/14), Compression Ratio=50%2026.02 | 78.9 | — | — | — | |
| AutoEncoder-VAEBackbone=CN-CLIP (ViT-H/14), Compression Ratio=75%2026.02 | 78.4 | — | — | — | |
| Q-FormerBackbone=CN-CLIP (ViT-H/14), Compression Ratio=75%2026.02 | 78.2 | — | — | — | |
| ARCBackbone=CN-CLIP (ViT-H/14), Compression Ratio=87.5%2026.02 | 78.2 | — | — | — | |
| AutoEncoder-VAEBackbone=CN-CLIP (ViT-H/14), Compression Ratio=87.5%2026.02 | 75.6 | — | — | — | |
| PTQ-P4QBackbone=CN-CLIP (ViT-H/14), Compression Ratio=87.5%, Bits=4-bit, Fine-tuning=true2026.02 | 75.1 | — | — | — | |
| Q-FormerBackbone=CN-CLIP (ViT-H/14), Compression Ratio=87.5%2026.02 | 74.8 | — | — | — | |
| PTQ-MinMaxBackbone=CN-CLIP (ViT-H/14), Compression Ratio=87.5%, Bits=4-bit, Fine-tuning=true2026.02 | 73.5 | — | — | — | |
| ARCBackbone=CN-CLIP (ViT-H/14), Compression Ratio=93.75%2026.02 | 73.3 | — | — | — | |
| AutoEncoder-VAEBackbone=CN-CLIP (ViT-H/14), Compression Ratio=93.75%2026.02 | 71.7 | — | — | — | |
| Q-FormerBackbone=CN-CLIP (ViT-H/14), Compression Ratio=93.75%2026.02 | 71.5 | — | — | — | |
| Pre-trained backboneProtocol=Zero-shot, Backbone=Qwen3-VL-8B, Resolution=224x224, Pooling Method=mean pooling2026.06 | 70.7 | — | 88.5 | 49 | |
| ViT-SBackbone=ViT, Embedding Dimension=3842024.05 | 47.8 | 60.2 | 72.2 | — | |
| ViT-SBackbone=ViT, Embedding Dimension=3842026.01 | 47.8 | 60.2 | 72.2 | — |