Data preparation:

July 21, 2025 ยท View on GitHub

Prepare the VG dataset and point the /path/to/image in the script to the VG dataset. The full dataset can be found at MM-GCoT.

Evaluation:

Execute evaluation scripts through the provided interface in ./scripts directory:

  • "SPLIT" refers to the dataset type, including "attributes", "judge", and "objects".
  • "CKPT" refers to the model name.
  • "answer_first" determines the generation order:
    • 1, the MLLM first generates the answer and then the bbox;
    • 0, the MLLM first generates the bbox and then the answer.