Visual Grounding on RCLIP-Bench Visual Grounding V1
96.5Success Rate (Att.)Human
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HumanScale=Unbounded Scale2026.06 | 96.5 | 95.1 | 95.3 | 97.1 | 94.7 | 95.7 | |
| GPT-4VScale=Unbounded Scale2026.06 | 92.7 | 83.6 | 78.2 | 85.5 | 94.5 | 86.5 | |
| ReasonSigLIPScale=ViT So400m (428M), Resolution=384/14, Data=+42M2026.06 | 32.1 | 38.5 | 17.2 | 25.1 | 41.1 | 30.8 | |
| ReasonCLIPScale=ViT Large (307M), Resolution=224/14, Data=+42M2026.06 | 25.8 | 36.7 | 15 | 19.3 | 36 | 26.6 | |
| ReasonCLIPScale=ViT Base (86M), Resolution=224/32, Data=+42M2026.06 | 22.5 | 31.4 | 14.2 | 18.5 | 31.4 | 23.6 |