deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
22.18k stars 5.28k forks source link

Dependencies list for face alignment #1101

Open yossibiton opened 4 years ago

yossibiton commented 4 years ago

I went over the README and couldn't find clear explanation on the face alignment process (which is a preprocessing for the training). I will assume this is a relevant script - "src/align/align_celeb.py". Am i right ?

However, this script has dependencies - the main one is MTCNN tensorflow version. Can you please mention the exact version of MTCNN you used (with url) ? There are many versions out there, i have seen your note : "Any public available MTCNN can be used to align the faces, and the performance should not change." But still it would be helpful to know which library you used, and if there are any other dependencies that should be installed (not mentioned at all at the README, except for mxnet)

bartolootrit commented 4 years ago

Hello. You may use https://github.com/deepinsight/insightface/tree/master/RetinaFace detector and https://github.com/deepinsight/insightface/blob/master/common/face_align.py for alignment

yossibiton commented 4 years ago

Thanks for the reference! But i see this detection is based on RetinaFace and not MTCNN. Which detector was eventually used to preprocess the training set ? Since I'm going to use the pretrained models I want to be sure the preprocessing pipeline stays the same

bartolootrit commented 4 years ago

You may use this MTCNN detector https://github.com/deepinsight/insightface/blob/master/deploy/mtcnn_detector.py. I'm not sure if it was used for training.

yossibiton commented 4 years ago

@bartolootrit thanks for the references ! But main question remains - what was the exact preprocessing used by this repository authors ? If we can't know that (e.g. exact face detector, exact alignment process) it's very hard the reproduce the LFW performance (99.8%)

oliversunde commented 2 years ago

Did you solve this yossibiton, there seems to be low documentation of how to use the face_align.py script, any inputs would be really helpful!