Image-to-Text Retrieval on MS-COCO
80.5R@1L2RM-SGRAF
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| L2RM-SGRAF2024.03 | 80.5 | 96.6 | 98.9 | 528.6 | — | |
| RCL-SGRAF2024.03 | 80.4 | 96.4 | 98.7 | 526.6 | — | |
| SGRAF2024.03 | 79.6 | 96.2 | 98.5 | 524.3 | — | |
| GSC-SGRNoise=20%2024.05 | 79.5 | 96.4 | 98.9 | 525.7 | — | |
| DECL-SGRAF2024.03 | 79.3 | 96.5 | 98.7 | 523.4 | — | |
| L2RM-SGR2024.03 | 79 | 96.4 | 98.3 | 523.4 | — | |
| BiCro-SGRAFNoise=20%2024.05 | 78.8 | 96.1 | 98.6 | 523.2 | — | |
| NCR2024.03 | 78.7 | 95.8 | 98.5 | 522.5 | — | |
| RCL-SAF2024.03 | 78.5 | 96.1 | 98.6 | 521.3 | — | |
| BiCro2024.03 | 78.4 | 95.6 | 98.5 | 520.5 | — | |
| RCL-SGR2024.03 | 78.2 | 96.2 | 98.4 | 521.4 | — | |
| L2RM-SAF2024.03 | 78.2 | 95.7 | 98.6 | 520.6 | — | |
| GSC-SGRNoise=40%2024.05 | 78.2 | 95.9 | 98.2 | 519.9 | — | |
| MSCN-SGRNoise=20%2024.05 | 78.1 | 97.2 | 98.8 | 524.6 | — | |
| SGR2024.03 | 78 | 95.8 | 98.2 | 518.1 | — | |
| DECL-SGRAFNoise=20%2024.05 | 77.5 | 95.9 | 98.4 | 518.2 | — | |
| BiCro-SGRAFNoise=40%2024.05 | 77 | 95.9 | 98.3 | 517.1 | — | |
| NCR-SGRNoise=20%2024.05 | 76.6 | 95.6 | 98.2 | 515 | — | |
| SAF2024.03 | 76.1 | 95.4 | 98.3 | 516.3 | — | |
| DECL-SGRAFNoise=40%2024.05 | 75.6 | 95.5 | 98.3 | 512 | — | |
| GSC-SGRNoise=60%2024.05 | 75.6 | 95.1 | 98 | 511.7 | — | |
| SGRAFNoise=20%2024.05 | 75.4 | 95.2 | 97.9 | 511.9 | — | |
| MSCN-SGRNoise=40%2024.05 | 75.3 | 95.4 | 98.2 | 512.6 | — | |
| NCR-SGRNoise=40%2024.05 | 74.7 | 94.6 | 98 | 509.7 | — | |
| IMRAM2024.03 | 74 | 95.6 | 98.4 | 512.1 | — | |
| BiCro-SGRAFNoise=60%2024.05 | 73.9 | 94.4 | 97.8 | 505.5 | — | |
| RINCE-SGRNoise=20%2024.05 | 73.8 | 95.6 | 98.5 | 513.5 | — | |
| DECL-SGRAFNoise=60%2024.05 | 73 | 94.2 | 97.9 | 502.5 | — | |
| MSCN-SGRNoise=60%2024.05 | 72.5 | 93.6 | 97.1 | 501.8 | — | |
| RINCE-SGRNoise=60%2024.05 | 72.3 | 94 | 97.9 | 501.7 | — | |
| FP32Backbone=SigLIP2-B/16, Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 71.6 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 71.54 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 71.54 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 71.52 | — | — | — | — | |
| RINCE-SGRNoise=40%2024.05 | 71.2 | 95.8 | 97.9 | 506.9 | — | |
| NoisyQuant w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 70.24 | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 70.12 | — | — | — | — | |
| ERQ w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 70 | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 69.76 | — | — | — | — | |
| ERQBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 69.76 | — | — | — | — | |
| PTQ4ViTBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 69.74 | — | — | — | — | |
| RepQ-ViTBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 69.2 | — | — | — | — | |
| MobileCLIP-BEvaluation Protocol=zero-shot2023.11 | 68.8 | 88.3 | 92.9 | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 67.7 | — | — | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 67.7 | — | — | — | — | |
| FIMA-QBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 67.68 | — | — | — | — | |
| ViCLIP-B/16Evaluation Protocol=zero-shot2023.11 | 67.2 | 87.3 | 92.7 | — | — | |
| ERQ w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 65.56 | — | — | — | — | |
| CLIP-ERMoE2025.11 | 65.4 | 88.3 | 94.7 | — | — | |
| CLIP-MoE2025.11 | 65 | 86 | 92 | — | — | |
| ERQBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 64.8 | — | — | — | — | |
| Naïve w/ RegCacheBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 64.02 | — | — | — | — | |
| MobileCLIP-S2Evaluation Protocol=zero-shot2023.11 | 63.4 | 85.1 | 91.4 | — | — | |
| Long-CLIP2025.11 | 62.8 | 85.1 | 91.2 | — | — | |
| Fix-CLIP2025.11 | 62.3 | 85.4 | 91.4 | — | — | |
| MobileCLIP-S1Evaluation Protocol=zero-shot2023.11 | 62.2 | 84.3 | 90.1 | — | — | |
| NoisyQuantBackbone=SigLIP2-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 62.04 | — | — | — | — | |
| FP32Backbone=OpenCLIP-B/16, Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 61.02 | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 60 | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 60 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 59.88 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 59.86 | — | — | — | — | |
| PTQ4ViTBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 59.6 | — | — | — | — | |
| ERQ w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 59.56 | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 59.44 | — | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 59.26 | — | — | — | — | |
| B-cosifiedMode=Zero-shot2025.12 | 59.1 | 81.1 | — | — | — | |
| OursMode=Zero-shot2025.12 | 58.9 | 80.9 | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 58.84 | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 58.78 | — | — | — | — | |
| ERQBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 58.74 | — | — | — | — | |
| MobileCLIP-S0Evaluation Protocol=zero-shot2023.11 | 58.7 | 81.1 | 88.2 | — | — | |
| CLIPMode=Zero-shot2025.12 | 58.7 | 80.7 | — | — | — | |
| RepQ-ViTBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 57.62 | — | — | — | — | |
| Naïve w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 57.6 | — | — | — | — | |
| FIMA-Q w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 57.34 | — | — | — | — | |
| RepQ-ViTBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 57 | — | — | — | — | |
| TinyCLIP-63M/32Evaluation Protocol=zero-shot2023.11 | 56.9 | — | — | — | — | |
| NoisyQuant w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 56.06 | — | — | — | — | |
| FIMA-QBackbone=OpenCLIP-B/16, Precision=W4A4, Evaluation Protocol=Zero-shot2025.10 | 55.82 | — | — | — | — | |
| PTQ4ViT w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 53.9 | — | — | — | — | |
| NoisyQuantBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 53.84 | — | — | — | — | |
| CLIPA-B/16Evaluation Protocol=zero-shot2023.11 | 53.1 | — | — | — | — | |
| NoisyQuant w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 51.58 | — | — | — | — | |
| NoisyQuant w/ RegCacheBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 50.66 | — | — | — | — | |
| PTQ4ViTBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 50.6 | — | — | — | — | |
| NoisyQuantBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 43.86 | — | — | — | — | |
| NaïveBackbone=OpenCLIP-B/16, Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 37.32 | — | — | — | — | |
| ERQ w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 33.64 | — | — | — | — | |
| RepQ-ViT w/ RegCacheBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 32.96 | — | — | — | — | |
| ERQBackbone=OpenCLIP-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 32.56 | — | — | — | — | |
| PTQ4ViTBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 29.88 | — | — | — | — | |
| NoisyQuantBackbone=SigLIP2-B/16, Precision=W6A6, Evaluation Protocol=Zero-shot2025.10 | 28.74 | — | — | — | — | |
| SGRNoise=20%2024.05 | 25.7 | 58.8 | 75.1 | 317.1 | — | |
| FLIPAttnmode=Zero-shot2024.05 | 23.89 | 48.33 | 60.04 | — | — | |
| Ours-Embeddingstrategy=Embedding, mode=Zero-shot2024.05 | 23.77 | 48.18 | 59.76 | — | — | |
| Ours-RGB0.3strategy=RGB, anchor_patch_ratio=0.3, mode=Zero-shot2024.05 | 23.65 | 47.54 | 59.25 | — | — | |
| CLIPmode=Zero-shot2024.05 | 23.49 | 47.8 | 59.66 | — | — | |
| Ours-RGB0.5strategy=RGB, anchor_patch_ratio=0.5, mode=Zero-shot2024.05 | 22.97 | 47.29 | 59.21 | — | — | |
| Ours-Kmeansstrategy=Kmeans, mode=Zero-shot2024.05 | 22.6 | 46.93 | 58.84 | — | — |