search

Gaoyiminggithub / Graphonomy

Graphonomy: Universal Human Parsing via Graph Transfer Learning
MIT License
295 stars 68 forks source link

How to obtain the "Category_rev_ids" images? #7

Closed zzw1123 closed 5 years ago

zzw1123 commented 5 years ago

Hi, Thanks for your contribution! I know that images in "Category_rev_ids" directory is obtained by flipping. But I also have to exchange the label of the body parts such as 'left-arm' and 'right-arm'? Am I right?

Gaoyiminggithub commented 5 years ago

Hi, You can download the 'rev' images. The link is in the README.md. The images downloaded from Readme.me in the 'rev' have been flipped which you don't need to exchange the label.

zzw1123 commented 5 years ago

@Gaoyiminggithub I have ignored that. Thanks for your kind reply!

zzw1123 commented 5 years ago

Sorry but I have another question...... In 'exp/transfer/train_cihp_from_pascal.py', there is a normalization applied to dataloader:

tr.Normalize_xception_tf(),

I do not know why you do this rather than standardizing them by: x-mean/std. Could you please tell me the difference between this two methods? Thanks a lot!

Gaoyiminggithub commented 5 years ago

Hi, It is an interesting question. Because the pretrained model is transferred from the Deeplab office code, the Deeplab code used the normalization way in here. So we use the same normalization way as it.

Gaoyiminggithub commented 5 years ago

Moreover, you can use the normalization way like x-mean/std to train the base model.

zzw1123 commented 5 years ago

I think the normalization method makes little effect to the final performance, right?