coin-dataset/code

code's Issues

Text in the video

Hi, Thanks for sharing the dataset!

I'm quite new to the dataset. After watching a few videos in the dataset, I noticed that some videos contain instructional text. For example: 8Cz14UJNUqE. Some even provide the action labels upfront, e.g. 8xawALD-cB4. Have you guys done any measures to prevent the possibility that the models pick up signals from the text instead of the content?

Background in transcripts for weakly supervised action segmentation?

Hi,
For the weakly supervised segmentation task, do you include background in the ordered transcripts?
For e.g., if a video has actions as (a_1, a_2, a_3) in respective order, but has background at the start, and in between a_2 and a_3, would the ordered transcript as per your paper's implementation be (b, a_1, a_2, b, a_3), or just (a_1, a_2, a_3)?
Thanks in advance!

Number classes used for step localization task

Hey,

For mAP calculation in Table 2 of the paper, how many classes was the evaluation done on? Its very unclear in the paper.

Trained SSN models

Thanks for the COIN dataset! Our group is working on improving performance on the COIN dataset, however, the training process takes quite some time, and having the trained models will help replicate your results exactly.
Is it possible to provide the trained SSN models for COIN?