Open jackspiderman opened 5 years ago
Sorry for the late response. I do realized an automatic landmark detection model (see "landmark regression" section in ReadMe). But it is trained on a very limited dataset and has poor accuracy when out of scope images are provided.
Of course, you can train that AAM model with you own data. Also, there are more advanced deep learning networks that achieve higher accuracy, but the bottleneck is always the training data.
I am attracted by this project. After reproducing it, I do feel the result is good. But I am a little bit sad that I have to generate 26 points manually for profile face. Is there any easier way to fix the trivial works problem? I would be very glad if anybody can give me some help.