skill_benchmark task

[sizhongzhang1]

Your work is excellent and I am very interested in your work. I found that the skill_benchmark task directly uses the video features extracted by I3D to perform the skill assessment task, may I know what I should do if I want to extract features from the original video and perform the skill assessment task

[hyf015]

Hi, thank you for your interest in this benchmark!

Yes, you can also use your own pretrained network to extract the video features. The video clips skill benchmark shows which frames are used in each video in their video names (see https://github.com/OpenGVLab/EgoExoLearn/tree/main/skill_benchmark).

You may either first trim the video using these frame numbers and then extract the clip-wise features, or first extract the whole video feature using a sliding window, and then locate the corresponding feature using the timestamp. There is no restriction in the feature extraction parameters (e.g., the size of the sliding window).