Question about the COCO-Object result

Start by reproducing the COCO-Object evaluation with the released implementation and the paper's settings, using the standard COCO-Object annotations and background-inclusive evaluation described in the issue. Compare the procedure with the reported Table 1 result, focusing on annotation conversion, prompts, dataset-specific parameters, post-processing, and the decoding bucket size; done means identifying and documenting the source of the 4.3 mIoU discrepancy or confirming the exact evaluation command and data-preparation procedure.

Hi, thank you for releasing the code.
I tried to reproduce the ActiveSAM result on COCO-Object using the released implementation and the settings provided in the paper. The results on VOC21 and PC60 are consistent with those reported in the paper, but the COCO-Object result is about 4.3 mIoU lower. I will attach my reproduction result below.
I used the standard COCO-Object annotations converted from COCO-Stuff-164K following the commonly used GroupViT/SCLIP protocol, and the evaluation includes the background class. I also verified that the decoding bucket size does not explain the difference.
Could you please clarify whether the reported COCO-Object result used different annotation files, conversion rules, prompts, dataset-specific parameters, or post-processing steps? If possible, could you also provide the exact evaluation command or data-preparation procedure used to obtain the result in Table 1?
Thank you for your help!