PointCNN parameter setting and other experience

Hi, Thanks for your sharing,it really helps.The ideas in this net are really great. I wonder which parameters I may set again if I want to use this net in a new situation.and what's more,can you share some experiences in the process of giving out this wonderful net?Thank you. I'm new hand in this field,forgive me for any unreasonable request.

Jul 22 '19 07:07 tangeroo

Hi @tangeroo Batch size, learning rate, xconv_params and fc_params is important. If you find that network is overfitting, just reduce the channel numbers in xconv_params or increase the dropout rate in FC layer. tuning sampling , epsilon or adding more information such as RGB is also work. Generally, if your task is classification, increase the receptive filed by tuning 'K', 'D' in xconv_params and segmentation task is opposite.

Thanks

Jul 23 '19 05:07 burui11087

Hi, I'm doing some classification on my own dataset and I've realised, when I change following settings: rotation_range = [0, math.pi, 0, 'u'] to rotation_range = [0, math.pi/18, 0, 'g'] I'm getting much better results. I've tried to understand (from code) what does it mean but unfortunately I wasn't successful. Could your please give me some hints or further links where I can find more information (or comments) about the settings parameter.

Thanks

Sep 11 '19 11:09 smgrz

@smgrz did you have any luck with the settings? I am getting overfitting and trying to see how i can make some of the changes

May 16 '20 08:05 dandanaus

If you find that network is overfitting, just reduce the channel numbers in xconv_params or increase the dropout rate in FC layer

@burui11087 can you please provide some more info on the segmentation side of things if we are seeing overfitting?

May 19 '20 10:05 dandanaus