understandings on this method

[lulujianjie]

Hi, Neeilan,

I watch the model code roughly. I think this method is just using a 3D-CNN classifier. That is, give some video as input and labels as output for CNN and force it to fit the dataset unreasonably.

Is my understanding correct?

[neeilan]

Hi @lulujianjie ,
The general idea is correct, but with a sufficiently large dataset trimmed to include only key events in a play, and by monitoring performance on an unrelated validation set, we can prevent overfitting.
I think that much of the error in predictions on the validation and test sets can be attributed to the quality of the downsampled videos used to limit model size.