Issue with "target_length" frames

[saliknadeem]

Hi,

Okay so I have been working with your code. Firstly I greatly appreciate your work, its very well written and easy to follow and understand. I have finally managed to run it and I am getting some results. There are somethings in the training process which I can't explain and I think I figured out one big issue.

For every action clip, you are using linspace to extract 6 frames from one entire action sequence. I think that is a big mistake! For example, in a sequence of 100 frames you are missing out 94% of the motion information!

I think we need to rewrite NTURGBDwithFlowLoader to get consecutive "target_length" frames from each video clip and use them all in the training process. Because currently you are only getting 6 frames which aren't even consecutive and we're using them to encode motion vector features.

[minostauros]

I really appreciate your kind feedback.
I'm currently away, but will be back in a couple of weeks to look into this issue.
Please feel free to share any additional ideas. Thanks.

[saliknadeem]

I found one other thing. It is not a major issue, but worth considering. In the paper, for reconstruction they are summing up all the other view reconstruction losses. In your code you are only selecting one other view randomly and using that loss. I modified the code to take both other views and sum the losses. But overall it doesnt seem to matter that much after a lot of epochs..