Reg. Training for face alignment

[mohanrajmit]

when i started running TrainDemo.cpp, it takes more than a day, what is the minimum number of images required for training and what is the minimum first_level_num ,second_level_num; in your code you have given first_level_num=10.

[sai-bi]

How many images do you have in training? It shouldn't take more than one day. If you run on the dataset provided my me, it will take about 3-4 hours in general.

There is no minimum first_level_num or second_level_num. But if you provide a too small number, the model got will perform poorly. I can only say that the parameters provided by me in the demo is enough to generate a good model on the provided dataset

[mohanrajmit]

1. The total number of images used for training is 1345. first_level_num=10, second_level_num=500. For each level minimum it requires more than one hour time for training.
2. when i reduced the first_level_num=2,second_level_num=10, the training finished very fast.

In the TestDemo.cpp line number 59 (cin>>index), what value should i give while running the code, when i gave index=1, i am getting error message.

[sai-bi]

1. I will try to improve the performance as soon as possible. You may check out and see the newest version 1 or 2 week later.
2. Of course, if you reduce second_level_num to 10 (original is 500), it will be much faster, however, the performance will be poor.
3. "index" should be the index of the test image. Make sure that that path to the test image is correct.

[mohanrajmit]

Thanks for your reply . now its working for me. suppose if i want to use my own database for training, then how to create the keypoints.txt, because i want to increase the number of landmarks.

[sai-bi]

keypoints.txt is in the following format: for each line, first list the x coordinate of each landmark, then list the y coordinate of each landmark, separated by space.

[mohanrajmit]

Manually we have to collect x and y coordinate of each landmark sir, or any other method to do..

is there any programming to do sir..

[sai-bi]

@mohanrajmit I have modified the program. It should be faster now, although it still cannot achieve the speed described by the paper :(

In terms of landmarks, of course you have to collect them manually if you want a new training dataset.

[mohanrajmit]

thanks for your immediate reply, where can i download the modified program for training. can you mail me the code to mohanraj4072@gmail.com, i can cite your paper.

[sai-bi]

Just download through github, which is the new version.

On Thu, Nov 13, 2014 at 11:29 AM, Mohanraj notifications@github.com wrote:

thanks for your immediate reply, where can i download the modified program for training. can you mail me the code to mohanraj4072@gmail.com, i can cite your paper.

— Reply to this email directly or view it on GitHub https://github.com/soundsilence/FaceAlignment/issues/3#issuecomment-62877533 .

毕 赛 / BI SAI

Department of Computer Science Faculty of Engineering University of Hong Kong Phone:(+852)54931076 Email:soundsilencebisai@gamil.com soundsilencebisai@gamil.com

[mohanrajmit]

thank u ..

[mohanrajmit]

what are the values must be present in boundingbox.txt file for training and testing.

Is it x,y,width and height of face region.