search

harryhan618 / SCNN_Pytorch

Pytorch implementation of "Spatial As Deep: Spatial CNN for Traffic Scene Understanding"
MIT License
250 stars 68 forks source link

why not use the std and mean values of CULane to do normalization when traing and testing? #61

Closed XavierNie715 closed 3 years ago

XavierNie715 commented 3 years ago

Thanks for your brilliant job. I'm little bit confused about the value the code used in normalization, why use std and mean values of imagenet to normalize?

harryhan618 commented 3 years ago

I think most probably because I'm using imagenet pretrained model weight. It won't cause much difference if you use customed dataset mean and std otherwise, according to my experience.

XavierNie715 commented 3 years ago

I think most probably because I'm using imagenet pretrained model weight. It won't cause much difference if you use customed dataset mean and std otherwise, according to my experience.

Thanks very much for your kindly response!