Mapping between 300 clips in Task 2/3 and 21 videos in Task 1

[adit98]

Hi,

I was wondering if the training/validation/testing splits of the 300 clips used in tasks 2 and 3 correspond to the train/val/test splits of the 21 videos used in task 1. In other words, are the 170 clips of the Task 2 training set all taken from the 10 training videos of task 1, and similarly for val and test? Thanks!

[ziyiwangx]

Hi @adit98

Thanks for your question.

The data splits of train/val/test for Task 2/3 correspond to the videos in Task 1 as follows: 1-10/11-16/17-21. Details about the clip-video data mapping information can be found in the supplementary material.

[adit98]

Ok, if I understood correctly:

Task 1 Train/Val/Test: 1-10 / 11-14 / 15-21 Tasks 2 and 3 Train/Val/Test: 1-10 (Clips 1-170) / 11-16 (clips 171-227) / 17-21 (clips 228-300)

Would it be possible to see the metadata on the mapping between the frames of each clip and the frames of the originating video?

[ziyiwangx]

Yes, the dataset splitting is the same as we used in the paper.

We did not record the precise timestamps for Task 2/3 corresponding to the Task 1 original videos. In the annotation file, we provide the mapping between each clip and the surgical phases in Task 1, which can be used as a rough reference if needed.