delphifirst
commented
9 years ago Hi Mu,
It's interesting to meet you on github.
The code cannot work for profile faces, because the training data I currently used are mostly frontal faces. I believe it can work for profile faces if we get suitable training data. However, I don't think frontal faces and profile faces can be handled in just one model, since their landmarks are very different ( some landmarks in profile faces are missing because of occlusion). Maybe the right way is first determining whether the face is frontal or profile, and then using the corresponding model.
The provided model can reproduce the accuracy reported by xudong's paper.
For CNN-based works, I have only read "Deep Convolutional Network Cascade for Facial Point Detection" in CVPR 2013. It says their method is slightly more accurate and much slower than xudong's method. I think CNN-based methods are interesting and practical, but it seems that they just applied CNN to face alignment problem and didn't reveal anything new for face alignment problem. ( Maybe this is unfair since I just read one paper in CNN-based workds)
I believe "Joint Cascade Face Detection and Alignment" in ECCV 2014 is the right direction. I think face detection, pose estimation and face alignment must be combined somehow. But I have no idea how to achieve this.
futurely
N1010N
hi yang,
many thanks for your codes, it is clean and works well.
but i tested on the new released casia-webface dataset, facex fails to generate meaningful landmarks on a large portion of images, mainly due to the failure of face detection. please see the figure below for an exmaple.
i wonder if there is any improvement space by using larger dataset or better algorithm tuning? and can the provided model reproduce the accuracy reported by xudong's paper? besides, there are also several cnn-based works, do you have any comment?
many thanks, mu