How to extract perframe features using mmaction2?

Hi,

I’m a beginner of video understanding. I have installed mmaction2 and denseflow, and it works well. But I’m not sure how to extract perframe features using it.

1. Is there any ready-made apis to solve the above problem? I’ve taken much time to search the solutions, but I didn’t find.
2. If no ready-made apis, could you plz share related codes with us? It’s important for beginners.
3. What is the difference between the config files eg. tsn_r50_320p_1x1x8_50e_activitynet_clip_rgb.py, tsn_r50_320p_1x1x8_50e_activitynet_video_rgb.py , specifically between video and clip?

Thank you for your awesome repo. And hope your reply.

[makecent]

By stating extrat perframe features, I am not sure what is the real case. But you may refer to this python file. https://github.com/open-mmlab/mmaction2/blob/fda023269e0fd9ed7ac36d48c67d2b86d422d602/tools/data/activitynet/tsn_feature_extraction.py#L1

[kennymckormick]

Yes, you can refer to mmaction2/tools/data/activitynet/tsn_feature_extraction.py for frame-level feature extraction. The difference between two configs is: clip_rgb uses clip-level annotations to train the recognizer (only use the groundtruth segments), while video_rgb uses video-level annotations (the whole untrimmed video).