search

fullfanta / CNN_based_Face_Alignment

Face Alignment using Convolutional Neural Networks
20 stars 11 forks source link

for the model #4

Open mathCrazyy opened 7 years ago

mathCrazyy commented 7 years ago

hello, thanks for you repo and i can get it for my mxnet order. and i want to know how about your trained model , is it have a good performance to closed eye, this is one point i attention,thank you ,and do you know more methods , i do not care about the speed , but the accuracy is important for me , thank you .

fullfanta commented 7 years ago

@mathCrazyy In public database - 300W which I used in this repository, there's no face image with closed eyes. To improve accuracy for closed eyes, you need to collect more data with closed eyes.

For the accuracy, the method I used here is not good. CNN estimates landmark locations with one feedforward. I recommend cascaded method such as MDM (Mnemonic Descent Method: A recurrent process applied for end-to-end face alignment", CVPR 2016 - https://github.com/trigeorgis/mdm). It constructs several cascades of CNN (or RNN with CNN).

mathCrazyy commented 7 years ago

thank you for your reply, and i just scan your advice MDM, and found this propgram is also used 300W dataset, umm..do you know some datasets include closed eye, after all , the Intraface have good performance when one person closed the eyes.

fullfanta commented 7 years ago

@mathCrazyy If you train SDM (method name for intraface) with 300W dataset, it will fails for closed eyes. I know that the published model file of intraface is trained with Youtube face dataset as mentioned in the paper.

You'd better to collect images with closed eyes.

mathCrazyy commented 7 years ago

well,thank you for your advice, and these days, i have trained intraface method and facealignment(https://github.com/1adrianb/face-alignment),and also dlib, indeed ,no one have a good performance for every case, i need to write a tool to correct the predicted points, and then use them to train a model, include the closed eye. The former good performance, i mean that , sometimes,the dlib will have some differeces in position with true landmark about eyebrow, also , the face alignment maybe have wrong position about the outline of face. in addition, the intraface is good almostly, but it only support 49 landmarks(no the outline of face),at the same time ,it also will be wrong ,but this is because the wrong position of face boundingbox,anyway.