search

ZhiwenShao / MCNet

ICME 2017 "Learning a Multi-Center Convolutional Network for Unconstrained Face Alignment"
16 stars 5 forks source link

The bounding box used to train COFW #1

Closed lsy1993311 closed 4 years ago

lsy1993311 commented 7 years ago

Hi, in your paper, you said you've process the data so each bounding box includes all landmarks, but unfortunately, the training bounding boxes provided by COFW are very bad.... So would you mind sharing the training bounding boxes used for training COFW?

ZhiwenShao commented 7 years ago

During training, for each face, we firstly perform multiple rotations, and obtain an initial face bounding box respectively according to the ground truth location of rotated landmarks. We next conduct scaling and translation on face bounding boxes with different extents, and crop faces on the basis of generated face bounding boxes. We finally augment training samples through horizontal flip and JPEG compression. It is noteworthy that we employ the provided face bounding boxes to crop face patches during testing.