Vision-language Compositional Reasoning on ARO Benchmark Visual Genome Flickr30k MS-COCO (test)
89.3VG AttributionCapPa
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CapPaArch=L/142023.06 | 89.3 | 86 | 99.3 | 99 | |
| CapArch=B/162023.06 | 88.9 | 86.6 | 99.1 | 99 | |
| ARO BestArch=-2023.06 | 88 | 73 | 60 | 46 | |
| CapPaArch=B/162023.06 | 85.7 | 86.7 | 99.2 | 98.8 | |
| Blind dec.Arch=-2023.06 | 83.7 | 86.2 | 98.8 | 98.7 | |
| NegCLIPArch=B/322023.06 | 71 | 81 | 91 | 86 | |
| CLIPArch=B/322023.06 | 63.2 | 59.1 | 59.4 | 47.3 | |
| CLIPArch=B/162023.06 | 62.7 | 58.7 | 57.9 | 49.5 | |
| CLIP* (16k)Arch=L/142023.06 | 57.8 | 35.9 | 40.2 | 31.5 | |
| CLIP* (8k)Arch=B/162023.06 | 55.4 | 39.8 | 43.7 | 32.8 | |
| CLIP* (16k)Arch=B/162023.06 | 53.2 | 39.7 | 45.5 | 37 |