Text-to-Image Retrieval on MS-COCO
67R@1BLIP (Cosine)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BLIP (Cosine)Backbone=BLIP, Alignment Strategy=Cosine2025.06 | 67 | 91.2 | 95.9 | — | — | — | |
| L2RM-SGRAF2024.03 | 65.7 | 90.8 | 96.1 | 528.6 | — | — | |
| GSC-SGRNoise=20%2024.05 | 64.4 | 90.6 | 95.9 | 525.7 | — | — | |
| RCL-SGRAF2024.03 | 64.3 | 90.8 | 96 | 526.6 | — | — | |
| MSCN-SGRNoise=20%2024.05 | 64.3 | 90.4 | 95.8 | 524.6 | — | — | |
| L2RM-SGR2024.03 | 63.7 | 90.2 | 95.8 | 523.4 | — | — | |
| BiCro-SGRAFNoise=20%2024.05 | 63.7 | 90.3 | 95.7 | 523.2 | — | — | |
| L2RM-SAF2024.03 | 63.4 | 89.6 | 95.1 | 520.6 | — | — | |
| GSMN (sparse+dense)Image Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 63.3 | 90.1 | 95.7 | 522.5 | — | — | |
| NCR2024.03 | 63.3 | 90.4 | 95.8 | 522.5 | — | — | |
| DECL-SGRAF2024.03 | 63.3 | 90.6 | 95 | 523.4 | — | — | |
| SGRAF2024.03 | 63.2 | 90.7 | 96.1 | 524.3 | — | — | |
| RCL-SGR2024.03 | 62.9 | 90 | 95.7 | 521.4 | — | — | |
| RCL-SAF2024.03 | 62.7 | 90 | 95.4 | 521.3 | — | — | |
| BiCro2024.03 | 62.6 | 89.7 | 95.7 | 520.5 | — | — | |
| GSC-SGRNoise=40%2024.05 | 62.5 | 89.7 | 95.4 | 519.9 | — | — | |
| SAF2024.03 | 61.8 | 89.4 | 95.3 | 516.3 | — | — | |
| BiCro-SGRAFNoise=40%2024.05 | 61.8 | 89.2 | 94.9 | 517.1 | — | — | |
| DECL-SGRAFNoise=20%2024.05 | 61.7 | 89.3 | 95.4 | 518.2 | — | — | |
| RINCE-SGRNoise=20%2024.05 | 61.7 | 89.2 | 94.7 | 513.5 | — | — | |
| PFANImage Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 61.6 | 89.6 | 95.2 | 518.2 | — | — | |
| SGR2024.03 | 61.4 | 89.3 | 95.4 | 518.1 | — | — | |
| NCR-SGRNoise=20%2024.05 | 60.8 | 88.8 | 95 | 515 | — | — | |
| IMRAM2024.03 | 60.6 | 88.9 | 94.6 | 512.1 | — | — | |
| GSMN (sparse)Image Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 60.4 | 88.7 | 95 | 514 | — | — | |
| GSMN (dense)Image Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 60.3 | 88.5 | 94.6 | 511.6 | — | — | |
| MSCN-SGRNoise=40%2024.05 | 60.3 | 88.6 | 94.8 | 512.6 | — | — | |
| SGRAFNoise=20%2024.05 | 60.1 | 88.5 | 94.8 | 511.9 | — | — | |
| GSC-SGRNoise=60%2024.05 | 60 | 88.3 | 94.6 | 511.7 | — | — | |
| NCR-SGRNoise=40%2024.05 | 59.6 | 88.1 | 94.7 | 509.7 | — | — | |
| DECL-SGRAFNoise=40%2024.05 | 59.5 | 88.3 | 94.8 | 512 | — | — | |
| BFANImage Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 59.4 | 88.4 | — | 317.9 | — | — | |
| RINCE-SGRNoise=40%2024.05 | 59.1 | 88.6 | 94.3 | 506.9 | — | — | |
| SCANImage Backbone=Faster R-CNN, Text Backbone=ft Bi-GRU2020.04 | 58.8 | 88.4 | 94.8 | 507.9 | — | — | |
| RINCE-SGRNoise=60%2024.05 | 58.4 | 86.6 | 92.5 | 501.7 | — | — | |
| BiCro-SGRAFNoise=60%2024.05 | 58.3 | 87.2 | 93.9 | 505.5 | — | — | |
| MSCN-SGRNoise=60%2024.05 | 57.7 | 87 | 93.9 | 501.8 | — | — | |
| DECL-SGRAFNoise=60%2024.05 | 57 | 86.6 | 93.8 | 502.5 | — | — | |
| SCOImage Backbone=fixed ResNet-152, Text Backbone=ft LSTM2020.04 | 56.7 | 87.5 | 94.8 | 499.3 | — | — | |
| GXNImage Backbone=ft ResNet-152, Text Backbone=ft GRU2020.04 | 56.6 | — | 94.5 | 317.5 | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 52.5 | — | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 52.49 | — | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 52.49 | — | — | — | — | — | |
| CARTParadigm=Coarse-Fine ID2024.06 | 52.42 | 77.53 | 86.11 | — | — | — | |
| FP32Backbone=SigLIP2-B/16, Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 52.33 | — | — | — | — | — | |
| VSE++Image Backbone=ft ResNet-152, Text Backbone=ft GRU2020.04 | 52 | — | 92 | 304.6 | — | — | |
| PTQ4ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 51.68 | — | — | — | — | — | |
| PTQ4ViTBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 51.62 | — | — | — | — | — | |
| NoisyQuant w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 51.41 | — | — | — | — | — | |
| MobileCLIP-BEvaluation Protocol=zero-shot2023.11 | 50.6 | 74.9 | 82.9 | — | — | — | |
| ERQ w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 50.2 | — | — | — | — | — | |
| RepQ-ViTBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 50.15 | — | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 50.15 | — | — | — | — | — | |
| ERQBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 49.97 | — | — | — | — | — | |
| CLIP (Cosine)Backbone=CLIP, Alignment Strategy=Cosine2025.06 | 49.3 | 76.4 | 88.2 | — | — | — | |
| Fix-CLIP2025.11 | 49.1 | 73.8 | 82.4 | — | — | — | |
| ViCLIP-B/16Evaluation Protocol=zero-shot2023.11 | 48.4 | 73.3 | 81.8 | — | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 47.22 | — | — | — | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 47.22 | — | — | — | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 47.21 | — | — | — | — | — | |
| DPCImage Backbone=ft ResNet-152, Text Backbone=ft ResNet-1522020.04 | 47.1 | 79.9 | 90 | 467.8 | — | — | |
| Naïve w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 46.8 | — | — | — | — | — | |
| CLIP-MoE2025.11 | 46.8 | 71.7 | 80.4 | — | — | — | |
| CLIP-ERMoE2025.11 | 46.5 | 72.9 | 81.6 | — | — | — | |
| NoisyQuantBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 46.32 | — | — | — | — | — | |
| Long-CLIP2025.11 | 46.3 | 70.8 | 79.8 | — | — | — | |
| ERQ w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 46.28 | — | — | — | — | — | |
| ERQBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 45.44 | — | — | — | — | — | |
| MobileCLIP-S2Evaluation Protocol=zero-shot2023.11 | 45.4 | 70.1 | 79 | — | — | — | |
| MobileCLIP-S1Evaluation Protocol=zero-shot2023.11 | 44 | 68.9 | 77.7 | — | — | — | |
| CLIPMode=Zero-shot2025.12 | 42.2 | 66.9 | — | — | — | — | |
| OursMode=Zero-shot2025.12 | 42.1 | 67.3 | — | — | — | — | |
| B-cosifiedMode=Zero-shot2025.12 | 41.9 | 66.7 | — | — | — | — | |
| GRACEParadigm=Atomic ID2024.06 | 41.5 | 69.1 | 79.1 | — | — | — | |
| FP32Backbone=OpenCLIP-B/16, Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 41.38 | — | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 41.26 | — | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 40.96 | — | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 40.96 | — | — | — | — | — | |
| PTQ4ViTBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 40.66 | — | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 40.58 | — | — | — | — | — | |
| MobileCLIP-S0Evaluation Protocol=zero-shot2023.11 | 40.4 | 66 | 75.9 | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 40.39 | — | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 40.3 | — | — | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 40.21 | — | — | — | — | — | |
| ERQ w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 40.04 | — | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 40.02 | — | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 39.96 | — | — | — | — | — | |
| ERQBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 39.83 | — | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 39.82 | — | — | — | — | — | |
| DSPEImage Backbone=fixed VGG-19, Text Backbone=w2v+HGLMM2020.04 | 39.6 | 75.2 | 86.9 | 420.7 | — | — | |
| NoisyQuant w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 39.25 | — | — | — | — | — | |
| RepQ-ViTBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 39.2 | — | — | — | — | — | |
| RepQ-ViTBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 38.88 | — | — | — | — | — | |
| CLIP (Manhattan)Backbone=CLIP, Alignment Strategy=Manhattan2025.06 | 38.8 | 68.7 | 80.1 | — | — | — | |
| TinyCLIP-63M/32Evaluation Protocol=zero-shot2023.11 | 38.5 | — | — | — | — | — | |
| Naïve w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 38.45 | — | — | — | — | — | |
| BLIP (Transformer MSE)Backbone=BLIP, Alignment Strategy=Transformer (MSE)2025.06 | 38.4 | 70.2 | 81.9 | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 38.02 | — | — | — | — | — | |
| Ours-RGB0.3strategy=RGB, anchor_patch_ratio=0.3, mode=Zero-shot2024.05 | 35.87 | 61.5 | 72.34 | — | — | — | |
| NoisyQuant w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 35.53 | — | — | — | — | — |