When do you plan to release the dataset and baselines?

[frostinassiky]

Hello,

Thanks for the amazing work!

I find this link from your paper "Video Question Answering with Phrases via Semantic Roles". When do you plan to release the dataset and baselines?

[TheShadow29]

@frostinassiky

My bad, reality has been catching upto me lately.

If you just need the files, here they are:

ASRL-QA:

Train: https://drive.google.com/file/d/1z5-kRyhqM_zTAEj0ONHF3qSzntyTVHG3/view?usp=sharing

Val / Test: https://drive.google.com/file/d/1D1Ja5WRzpq6tL6RhE6TD0NpcA9UCIYwZ/view?usp=sharing

Charades-SRL-QA:

Train: https://drive.google.com/file/d/1-83jwIEsbmTI3_aT0wVBFlNk_ZtnaqCa/view?usp=sharing

Val / Test: https://drive.google.com/file/d/1tNMSsp5AaTu7qMVDAoNqbIg3-tQ7mW8j/view?usp=sharing

The way to read the files is as follows:

1. Each file is a list of dictionaries. Each dictionary has many keys, but the required ones are:

• vid_seg: The relevant video segment name
• qa_pair: This points to a dictionary with the keys question and answer. If you just want to see the question/answers this all you would need.
• qsrl_ind: a unique identifier for the annotation
• cs_qsrl_inds: a list of indices pointing to contrastive samples. These are only useful for validation / testing
• rand_qsrl_inds: similar to above, but some random ones (not contrastive). Useful for validation / testing.

I will try to push the changes with data+baselines at the latest. Please expect it before June 3rd.

[frostinassiky]

Thanks for the quick response! I will follow it up on June 3rd. :p