Evaluation setup for the activity classification task

Hi, thank you very much for the excellent work and for open sourcing your project!

I'm trying to reproduce some of your results, particularly the activity classification accuracies and I was wondering how is one sample defined in that downstream task.

In other words, is it:
1) fixed-duration crop from the whole take associated with the take's global activity label or
2) sample associated with atomic action description with certain context window around it associated with the label of the take to which it belongs?

If it is the first option, how do you select that crop? Is scenario_cls.csv used for this task or?

I would appreciate any guidance. Thanks a lot!

Start by tracing the activity-classification evaluation setup and inspect how scenario_cls.csv is used. Determine whether samples are fixed-duration crops or action-centered windows, how crops are selected, and which label is assigned. Done means the sampling and labeling procedure is documented clearly enough to reproduce the reported accuracies.

Written by the indexing model from the issue text.