How to predict without face detection?

[JIEMIN1995]

I'm kind of confused in the testing stage. The paper declares the DAN can still obtain good results without face detections but I cannot find implementation details about this step. I tried to use a fixed small bbox in the center of picture to initialize the mean shape while this method failed. Anybody could give me suggestions? Thanks a lot.

[MarekKowalski]

Hi,
In the paper DAN is tested on a two datasets 300-W and Menpo. The Menpo dataset does not have face detection bounding boxes, so for this dateset I decided to initialize DAN with a bounding box located in the center of the image. Keep in mind that this only worked because in the images in that dataset all of the faces are located approximately in the center of the image.

If you initialize the bounding box in a location that is close to the true face location and at a close scale, it should work reasonably. If you initialize it far away from the face or at a vastly different scale, it won't.

Thanks,

marek