search

juntang-zhuang / ShelfNet

implementation for paper "ShelfNet for fast semantic segmentation"
MIT License
252 stars 41 forks source link

About data augmentation #9

Closed kaixinxiao closed 5 years ago

kaixinxiao commented 5 years ago

I find a problem in train.py: # data transforms input_transform = transform.Compose([ transform.ToTensor(), transform.Normalize([.485, .456, .406], [.229, .224, .225])]) It seems different from paper saying:

For data augmentation, the image is randomly flipped and scaled between 0.5 to 2, and randomly rotated between -10 and 10 degrees. Should I add some data augmentation methods like paper's description? Thank you.

juntang-zhuang commented 5 years ago

Thanks for your interest. The data augmentation parameter parsing is not well written, some part of it is written within the dataset class. But our results in the old paper are generated from the same version in this repo. Besides, we have some updates in the code and updated results in our new version paper, but neither the new code nor the new paper can be published now., will release them in the future.