Evaluation on AVVP dataset

[PREPONDERANCE]

This is taken from crab/dataset/unified_dataset.py file. I wonder where I can find the data/avvp_data/train_samples.json file. Another question is that since AVVP dataset provides fixed event labels, how to evaluate the model's open-ended prediction against the fixed label ground truth?

[CserDu]

@PREPONDERANCE Hello, the AVVP train_samples.json file has been uploaded to the path: data/train_samples.json
For evaluation, the regular expression is leveraged to parse the structured results of the model's predictions. For example, <visual>Dog,(0 10)</visual> <audio>Speech,(0 10)</audio>.

[PREPONDERANCE]

Thanks for your reply! I've also found the relevant files in the Huggingface dataset repo. I have another question though. Since the paper claims sourcing the AVSBench dataset to construct the ARIG task, I wonder what it means to write 0GnvsN_8_Sk_4000_9000 as the filename? Especially, the 4000-9000 number digits are not found in the original dataset.

[PREPONDERANCE]

Also, it seems that the v1s and v2 grounding annotations sourcs from Ref-AVS as well. Would you elaborate the exact source datasets and describe how you reindex (rename) those files?