3D convolution for Deep Neural Network (DNN) module

asked 2018-11-09 02:29:13 -0500

l.fabbrini gravatar image

Hi all, Is there any plan to support network models which have 3D convolution and 3D Max Pooling? I mean the input to the network is 5D (e.g. "NDHWC" that is: batch-depth-height-width-channel). At now (OpenCV 4.0.0-beta) readNet() function only supports "NHWC" and "NCHW" data format.

I think the support of 3D convolution and 3D Max Pooling would be very important for the community, a lot of volume data (Video, Medical Images, etc.) are processed with this type of CNN. Thanks, L

edit retag flag offensive close merge delete



From my experience it's still the best to use the original framework for DNN. I think it's much faster (on GPU) and you avoid these conversion, adaptation and "unknown layer type" problems...

kbarni gravatar imagekbarni ( 2018-11-09 07:15:35 -0500 )edit

I agree - i initially planned to use open cv as main api across all the cnn. I noticed that this is not a good idea because of performance and compatibility issues. I am using a micro service architecture to get rid of these problems and its working fine :-)

holger gravatar imageholger ( 2018-11-09 07:49:14 -0500 )edit