search

noureldien / timeception

Timeception for Complex Action Recognition, CVPR 2019 (Oral Presentation)
https://noureldien.com/research/timeception/
GNU General Public License v3.0
157 stars 33 forks source link

Question regarding to Conv2D #4

Closed sudonto closed 5 years ago

sudonto commented 5 years ago

Hi @noureldien , nice work. In your paper, you put Conv2D in Temporal Conv Module. But, after plotting your model to jpeg, I did not see the presence of Conv2D layer in that module.

image

In my graph, it's not Conv2D but BatchNorm layer. Can you explain more about this? Thanks.

noureldien commented 5 years ago

Hi, thanks for asking. Conv2D here means 1x1x1 Conv3D dimension reduction. You find it here:

Keras/Tensorflow https://github.com/noureldien/timeception/blob/master/nets/timeception.py#L120 Either you do depth-wise conv (Cov1D) then dimension reduction (1x1x1 Conv3D), or you do first dimension reduction (1x1 Conv2D) followed by depth-wise conv (Cov1D). Both will get you the same result.