TheShadow29 / VidSitu

[CVPR21] Visual Semantic Role Labeling for Video Understanding (https://arxiv.org/abs/2104.00990)
https://vidsitu.org/
MIT License
56 stars 8 forks source link

About the 'vb_lst' for vinfo_files for validation #27

Closed Versocial closed 8 months ago

Versocial commented 9 months ago

in data/README.MD:

vinfo_files/:
...
For Validation file, the following keys are additionally present:

vb_to_use: The verbs used to gather Semantic Roles and Event Relations.
vb_lst: A Dictionary with keys Ev1-Ev5 and values as a list of 10-verbs
vbid_lst: Same as above, but instead of verb names, uses the verb ids.

I know the 'vb_lst' is the candidates for 'vb_to_use', but what is the 'vb_lst' used for ? Or is there a 'vb_lst' for the Training data too ? Maybe the candidates can be used when training for verb prediction.

TheShadow29 commented 8 months ago

@Versocial Sorry for the later reply.

The vb_lst is because the validation datasets have 10 verbs for each event. During training, we ask the annotator to directly produce a verb+srl+event relation.

For validation, we ask 10 annotators only the verb followed by top-consensus verb to have 3 SRL annotations.

Versocial commented 8 months ago

Quite Clear. Thanks for your answering . I'll close this issue.