search

paulu / deepfeatinterp

Deep Feature Interpolation (CVPR 2017)
GNU General Public License v3.0
278 stars 50 forks source link

LFW Preprocessing details. #4

Open black-puppydog opened 7 years ago

black-puppydog commented 7 years ago

In the paper, you write that you use dlib to preprocess data.

Given that the images in this file: https://github.com/paulu/deepfeatinterp/blob/76b72e0a3b683a8cdeb13ed35b932680dc159ed1/demo1.py#L75 all differ from the originals, and from all the aligned datasets provided on the LFW website, could you elaborate on how it was created? The filename seems to imply a connection with AEGAN, but the aegan repo doesn't use dlib at all. So I'm not sure how to go from the original lfw dataset to the one you used, and I don't find this in this repo either, at least not for the data used in demo1.py.

Oh and by the way, the file contains 13238 .jpg files (and some extra ._xxx.jpg files that OS X put in there). The five extra images are Ralf_Schumacher_0005_* images that are actually results, not inputs.

paulu commented 7 years ago

Dlib is not used in demo1. We use the same alignment as AEGAN in order to control for that factor in our comparison. Specifically, this means we used the images from DeepPy. The images in lfw_aegan.tar.gz were created by running AEGAN and extracting the individual images from the cache files generated by DeepPy.

Thanks for pointing out the extra files.