Image Retrieval on MS COCO Karpathy's split (val)
12.5Deletion @1Grad-ECLIP
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Grad-ECLIPBackbone=CLIP2025.02 | 12.5 | 26.7 | 15.8 | 32 | 0.033 | 0.051 | — | — | — | |
| MaskCLIPBackbone=CLIP2025.02 | 13.2 | 28.4 | 14.2 | 29.5 | 0.01 | 0.038 | — | — | — | |
| Grad-ECLIP w/o λiBackbone=CLIP2025.02 | 13.9 | 29.4 | 14 | 28.9 | 0.001 | -0.009 | — | — | — | |
| raw attentionBackbone=CLIP2025.02 | 17.1 | 35.5 | 12.5 | 25.5 | -0.046 | -0.038 | — | — | — | |
| GAMEBackbone=CLIP2025.02 | 17.1 | 35.5 | 15.4 | 30.8 | -0.017 | 0.012 | — | — | — | |
| Grad-CAMBackbone=CLIP2025.02 | 17.2 | 35 | 10.3 | 22.2 | -0.069 | -0.101 | — | — | — | |
| CLIPSurgeryBackbone=CLIP2025.02 | 17.9 | 36.5 | 14.2 | 29.4 | -0.037 | -0.061 | — | — | — | |
| M2IBBackbone=CLIP2025.02 | 17.9 | 36.7 | 14.7 | 30 | -0.033 | 0 | — | — | — | |
| RolloutBackbone=CLIP2025.02 | 19.5 | 39.5 | 12.9 | 29.3 | -0.065 | -0.052 | — | — | — | |
| GAME2025.02 | — | — | — | — | — | — | 10.8 | 11.5 | 0.006 | |
| Grad-ECLIPlambda_i=included2025.02 | — | — | — | — | — | — | 10 | 12.9 | 0.076 | |
| Grad-ECLIPlambda_i=excluded2025.02 | — | — | — | — | — | — | 11.2 | 11.2 | 0.001 | |
| M2IB2025.02 | — | — | — | — | — | — | 21.4 | 0.5 | -0.233 | |
| raw attention2025.02 | — | — | — | — | — | — | 28.4 | 0.7 | -0.278 | |
| Rollout2025.02 | — | — | — | — | — | — | 12.2 | 10.5 | -0.017 |