zhenglab/spiralnet

test masks

Opened this issue · 1 comments

Does the masks made from random squares?
At least I found that the bboxes in celebA are not 128x128 -> 256x256 even after resizing.
or could you share the masks?

The masks are square and generated randomly (128128 white region in 256256 black background), and we just use same random masks for the comparison of different methods.

We use CelebA-HQ dataset rather than CelebA, whose images are square.
I think you may crop out the bboxes region of CelebA faces and resize to 256*256, and randomly generate square masks for training and enough square masks for testing. Thanks.