search

srijandas07 / VPN

Pose driven attention mechanism
37 stars 9 forks source link

Data Normalization #5

Closed tranminhduc4796 closed 3 years ago

tranminhduc4796 commented 3 years ago

Hi, I am trying to train NTU with your model but your code does not mention how the poses' x, y, z are normalized. Could you give me your normalization method.

srijandas07 commented 3 years ago

@tranminhduc4796 a very good data normalization of the poses is performed in https://github.com/lshiwjx/2s-AGCN In our case, we just perform a body translation with the hips (already in the Loader). We do not perform any other normalization.

tranminhduc4796 commented 3 years ago

Thanks. I also have another suspicious and hope you can confirm it for me. A npy file you mentioned contain a matrix represent 2 poses in a single video so each has a size of n_frames x 150 (25 joints x 3 coords x 2 people). Is my understanding right?

srijandas07 commented 3 years ago

Yes, absolutely the matrix represents 2 poses. In case, there is a single person in a video, we pad the last 75 elements with zeros.

tranminhduc4796 commented 3 years ago

Thank you for all of your supports.