Scene Graph Generation on PVSG IoU threshold = 50 (test)
1.068Generation Time (s)SceneGraphVLM (SFT)
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SceneGraphVLM (SFT)Prompting=GT prompt, Train split=PSFR2026.05 | 1.068 | 0 | 72.5 | 72.7 | 72.5 | 72.7 | 72 | 72.2 | 63.5 | 63.6 | 63.1 | 63.2 | 62.2 | 62.4 | 67.1 | 67.3 | |
| SceneGraphVLM (SFT+RL)Prompting=GT prompt, Train split=BaseAnnot2026.05 | 1.068 | 0 | 77.9 | 78.1 | 73.1 | 73.3 | 74.8 | 75 | 67.3 | 67.4 | 61.4 | 61.5 | 62.8 | 62.9 | 68.8 | 69 | |
| SceneGraphVLM (SFT)Prompting=Generated prompt, Train split=BaseAnnot2026.05 | 1.068 | 0 | 50.8 | 53 | 42.3 | 44 | 45 | 46.9 | 19.4 | 21.7 | 15.8 | 18.6 | 16.3 | 18.5 | 30.6 | 32.7 | |
| SceneGraphVLM (SFT+RL)Prompting=Generated prompt, Train split=BaseAnnot2026.05 | 1.068 | 0 | 57.4 | 59.9 | 46.9 | 49.1 | 50.4 | 52.8 | 29.5 | 31.9 | 20.2 | 22.2 | 22.7 | 24.7 | 36.5 | 38.8 | |
| GPT-5.4-miniPrompting=Generated prompt2026.05 | 2.4 | 77 | 11.6 | 30 | 8.9 | 21.4 | 9.6 | 23.7 | 4.3 | 5.8 | 2.5 | 3.4 | 2.7 | 3.8 | 6.2 | 13.7 | |
| GPT-5.4-nanoPrompting=GT prompt2026.05 | 2.678 | 2.23 | 14.4 | 20.2 | 12.9 | 18.3 | 12.9 | 18.1 | 1.9 | 2.2 | 2.1 | 2.2 | 1.9 | 2 | 7.4 | 10.1 |