grounding weak

[Troy0630]

For the grounding task, does MedMO output coordinates normalized to 0–1000 like Qwen3-VL, or does it use absolute image dimensions?

Also, why is the performance of MedMO-Next4B on MedSG-Bench very poor in my experiments, far below the results reported in the paper?

If possible, could you please share the prompt you used for this task?

this is normalized to 0–1000 result table：

task	samples	mean IoU	IoU@0.1 recall	IoU@0.3 recall	IoU@0.5 recall	valid parse
multi_view	1148	0.1118646637	0.2177700348	0.1611498258	0.1097560976	0.3928571429
object_tracking	1000	0.0899920427	0.2000000000	0.1270000000	0.0760000000	0.3360000000
referring	1482	0.2232926066

task samples mean IoU IoU@0.1 recall IoU@0.3 recall IoU@0.5 recall valid parse
multi_view 1148 0.1118646637 0.2177700348 0.1611498258 0.1097560976 0.3928571429
object_tracking 1000 0.0899920427 0.2000000000 0.1270000000 0.0760000000 0.3360000000
referring 1482 0.2232926066

[ankan8145]

Hi, for MedSG-Bench, you don’t need to normalize coordinates to 0–1000. MedMO uses absolute image coordinates.

The low MedMO-Next4B result may be due to using normalized coordinates during evaluation. Please rerun with absolute coordinates, and benchmark both MedMO-Next4B and MedMO-Next8B as mentioned in the MedSG-Bench GitHub.