Text Retrieval on MS COCO Karpathy's split (val)
23.8Deletion @1CLIPSurgery
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| CLIPSurgeryBackbone=CLIP2025.02 | 23.8 | 42.9 | 17.7 | 33.8 | -0.071 | -0.091 | — | — | — | |
| RolloutBackbone=CLIP2025.02 | 22.7 | 42.4 | 17.5 | 35 | -0.128 | -0.105 | — | — | — | |
| Grad-CAMBackbone=CLIP2025.02 | 21.6 | 40.1 | 11.5 | 23.3 | -0.129 | -0.168 | — | — | — | |
| M2IBBackbone=CLIP2025.02 | 20.6 | 39.1 | 20.6 | 36.9 | -0.067 | -0.021 | — | — | — | |
| GAMEBackbone=CLIP2025.02 | 19.8 | 38 | 21 | 37.4 | -0.047 | -0.006 | — | — | — | |
| raw attentionBackbone=CLIP2025.02 | 19.2 | 37.2 | 15.4 | 29.7 | -0.1 | -0.075 | — | — | — | |
| Grad-ECLIP w/o λiBackbone=CLIP2025.02 | 18.3 | 33.9 | 17.4 | 32.8 | -0.005 | -0.011 | — | — | — | |
| Grad-ECLIPBackbone=CLIP2025.02 | 15.5 | 29.3 | 20.6 | 37.6 | 0.053 | 0.083 | — | — | — | |
| MaskCLIPBackbone=CLIP2025.02 | 15.2 | 29.5 | 18.9 | 35.1 | 0.011 | 0.057 | — | — | — | |
| GAME2025.02 | — | — | — | — | — | — | 20.8 | 23 | 0.022 | |
| Grad-ECLIPlambda_i=included2025.02 | — | — | — | — | — | — | 17.7 | 25.4 | 0.03 | |
| Grad-ECLIPlambda_i=excluded2025.02 | — | — | — | — | — | — | 21.1 | 23.6 | 0.025 | |
| M2IB2025.02 | — | — | — | — | — | — | 42.6 | 2.4 | -0.419 | |
| raw attention2025.02 | — | — | — | — | — | — | 49.2 | 3.3 | -0.459 | |
| Rollout2025.02 | — | — | — | — | — | — | 23.9 | 20.7 | -0.032 |