search

davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.79k stars 4.81k forks source link

Why do the the face thumbnails crop to 182*182 pixels while using MTCNN? #738

Open wangyujie413 opened 6 years ago

wangyujie413 commented 6 years ago

When I crop face thumbnails to 182182 pixels,the systerm performs better than when crop to 160160 directly,so why? How to get the size of 182?

Dyex719 commented 6 years ago

I guess that the feature vectors are generated better maybe because of the increase in the amount of information. 182 is the default parameter, you can change it according to your wish. Look at: https://github.com/davidsandberg/facenet/blob/master/src/align/align_dataset_mtcnn.py#L147

Edit: The images which are 182x182 pixels are later converted to 160x160 pixels when training. This enables random cropping of the image, which prevents over-fitting.