Affordance Grounding on A2A-Bench Multi-instance 1.0 (test)
60.44sIoUA2A-AffordGen
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| A2A-AffordGenGrounding source=Iterative points2026.06 | 60.44 | 72.05 | 70.56 | 80.17 | 57.63 | |
| SAM3+textGrounding source=Text only2026.06 | 46.82 | 56.76 | 66.33 | 60.32 | 46.03 | |
| LISA-7B (finetune)Grounding source=MLLM mask, Protocol=finetune2026.06 | 40.26 | 56.38 | 64.23 | 58.93 | 37.14 | |
| Qwen3-VL-32B-Instruct+SAM3Grounding source=VLM box2026.06 | 29.45 | 35.99 | 57.78 | 38.1 | 21.75 | |
| GroundingDINO+SAM3Grounding source=Open-vocab. box2026.06 | 28.23 | 36.13 | 25.06 | 28.57 | 17.94 | |
| LISA-7BGrounding source=MLLM mask2026.06 | 20.05 | 29.37 | 28.73 | 23.81 | 15.24 | |
| Qwen3.5-9B+SAM3Grounding source=VLM box2026.06 | 18.55 | 22.38 | 29.34 | 19.05 | 7.14 | |
| Segagent (Qwen-VL-7B+SAM)Grounding source=Iterative points2026.06 | 18.51 | 26.23 | 21.84 | 15.87 | 8.41 |