The code of "Training-Free Confidence Scoring Mechanism"

[CongHan0808]

Thanks for sharing your work. Are there some codes for "Training-Free Confidence Scoring Mechanism"? I clone the repository and only find eval/run_llava.py for runing a demo. And are there the evaluation code for coco?

[jefferyZhan]

Hi @CongHan0808 , Thank you for your interest in our work. In the init version of this repo, we only provided the demo code. We'll release the training and batch inference code later. Before the release, you can follow the instruction to modify the demo code:

1. Set the return_dict_in_generate=True and output_scores=True in the model.generate function.
2. Get the output sequence and corresponding score of each output token.
3. Calculate the bbox score and category score of each instance according to the paper (our implementation is the same as the paper description).

Later, we'll release the whole codes including the batch inference of REC/REG/Detection/Counting/Phrase Grounding.