search

Seanlinx / mtcnn

594 stars 264 forks source link

facial landmark regression #37

Open Rainy000 opened 7 years ago

Rainy000 commented 7 years ago

Hello@Seanlinx,According to issue I finished facial landmark based on your Onet, the location results is not very robust. It looks very well on some picture but sometimes the deviation is big. How to adjust the super parameters to optimize the result? PS: I already try some methods such as : learning rate(adjust to 0.001)/dropout(0.25)、weight decay(more bigger)..

Seanlinx commented 7 years ago

@Rainy000 Try clip_gradient while initializing optimizer in train.py

Rainy000 commented 7 years ago

@Seanlinx I will try it, thank you very much!

Seanlinx commented 7 years ago

@Rainy000 Sorry, I mistook deviation for derivative, so clip gradient is not a solution. Did you follow the ratio the author mentioned in the paper? I think raising the importance of landmark alignment task will probably help.

Rainy000 commented 7 years ago

@Seanlinx Of course, but I just do face alignment on Onet, and the ratio of tasks is: face detection : 1、box regression: 0.5、facial landmark: 1. The paper do face alignment in three phases(Pnet Rnet Onet), but I do not know how to transfer the location results of Pnet and Rnet to Onet. Can you give me some suggestions?

Seanlinx commented 7 years ago

@Rainy000 You can test the celebA dataset using your Rnet model and compute normalized error of predicted landmarks, then choose those with big deviations to be the training set of Onet

Rainy000 commented 7 years ago

@Seanlinx It is similar to the hard examples of face detection training set?

Seanlinx commented 7 years ago

@Rainy000 Exactly.

pribadihcr commented 7 years ago

@Rainy000,

I want to fine tune the landmark also. Some times MTCNN bad in predict the landmark point in suppose 90 degree rotation face. Can you give me step by step how to do it. Thanks

Rainy000 commented 7 years ago

@pribadihcr I have not tested the landmark result on 90 degree rotation face. If you want to train the landmark model, you can refer to https://github.com/Seanlinx/mtcnn/issues/25

HaoLiuHust commented 6 years ago

@Rainy000 Can you share your implementation?

Rainy000 commented 6 years ago

@HaoLiuHust Of course. I just implement the facial landmark based on the codes of @Seanlinx , if you are interested in this , I will take some time to look for this project. And how to give you?

HaoLiuHust commented 6 years ago

@Rainy000 Appreciate your help, you can give me a link or send to my email: haoliuhust@hotmail.com , thank you again

Rainy000 commented 6 years ago

@HaoLiuHust I have sent the codes package to your email , if you have some questions, we can discuss together, good luck~

HaoLiuHust commented 6 years ago

@Rainy000 Thank you very much, but seems I did not receive the email, could you have a look?

HaoLiuHust commented 6 years ago

@Rainy000 Sorry, found it

xiaoxiongli commented 6 years ago

Dear Rainy000, @Rainy000

Could you please send me your code, i train a model base on Seanlinx's code, And i also need add face landmark detection. Please send me and let's discuss it. my email: 19295656@qq.com

thank you very much! ^_^

Rainy000 commented 6 years ago

@xiaoxiongli Sorry, just see your message. My company does not have network, I will send you when I go home~

xiaoxiongli commented 6 years ago

@Rainy000 it is very kind of you～ thank you

linsonwang commented 6 years ago

hi @Rainy000, could you please send me a copy of face landmark detection? Many thanks!

Rainy000 commented 6 years ago

@linsonwang give me your email~

Mumu16 commented 6 years ago

@Rainy000 hi Rainy000, could you please send me a copy of facial landmark detection? My email: jing7994@163.com. Thanks in advance!

zzzzzz0407 commented 6 years ago

As your code is popular, i think you can share you code on the github.

KiXiang commented 6 years ago

@Rainy000 hi Rainy000, could you please send me a copy of facial landmark detection? My email:809209272@qq.com

Rainy000 commented 6 years ago

@KiXiang @Mumu16 OK, GOOD LUCK~

Rainy000 commented 6 years ago

@zHanami Good ideas, but I am not familiar with how to upload code. The most important key is that the code is implemented based on @Seanlinx , so it's not good if I upload to my account.

mrgjbd commented 6 years ago

@Rainy000 hello Rainy000,I get stuck on Landmarks, could you please send me a copy of facial landmark detection? Thanks a lot，email：494587502@qq.com

OceanWong1991 commented 6 years ago

@Rainy000 Hi Rainy000! recently, I rewrite MTCNN by TensorFlow, but the Landmark predict is not precise, can you share your code with me, Thanks very much! email : OceanWong1990@gmail.com

Rainy000 commented 6 years ago

@OceanWong1991 Sorry, I quit from last company and have not backup copy of this project. You can ask for help to upstairs user, or implement it refer to issue #25.

tzhang2014 commented 6 years ago

@Rainy000 can you share your code for training landmark with me ? thanks ! email :416734877@qq.com

Fei-dong commented 6 years ago

@Rainy000 can you share your code for training landmark with me ? thanks ! email :1158563915@qq.com

Rainy000 commented 6 years ago

@githublet Try to expand your trainset, and keep the OHEM turning on. Maybe, you can try some other percentage.

Rainy000 commented 6 years ago

@githublet If you train face detection and landmark jointly, the first label ( -1 ) represent part example which with IOU between [0.45,0.6], and used to box regression. you can choose other value ( such as -2 ) represent landmark examples . So the train list could be below: img_path -2 x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 Of course, the landmark label used to jointly training could be below img_path -2 0 0 0 0 x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 and the value of x,y belong to (0,1) landmark training is regression problem, it hasn't neg example. OHEM trick can impose the robustness in complex scene. BTW, witch->which?

Rainy000 commented 6 years ago

@githublet From scratch. I know -1 mean ignore label, but you have to know which img is landmark sample. When training face detection, you can ignore part samples and landmark samples with label -1, but when training landmark ,you have to pick out the landmark samples with a distinguished label. I am chinese of course.

Rainy000 commented 6 years ago

@githublet The reason why give landmark samples a distinguished label is to training cls, box regression, landmark regression together. You should look this project more carefully.

PhenomenalOnee commented 5 years ago

@Rainy000
can you provide your code and data for training the three nets seperately for both face landmark and detection

cuiyong127 commented 5 years ago

@Rainy000 hello Rainy000,I get stuck on Landmarks, can you share your code for training landmark with me ? thanks ! email :359894979@qq.com

gm19900510 commented 4 years ago

hi Rainy000, could you please send me a copy of facial landmark detection? My email:1025304567@qq.com

cuiyong127 commented 4 years ago

hi,i also need the code of facial landmark detection.

------------------ 原始邮件 ------------------ 发件人: "gm19900510"<notifications@github.com>; 发送时间: 2020年1月2日(星期四) 下午3:28 收件人: "Seanlinx/mtcnn"<mtcnn@noreply.github.com>; 抄送: "Richar"<359894979@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [Seanlinx/mtcnn] facial landmark regression (#37)

hi Rainy000, could you please send me a copy of facial landmark detection? My email:1025304567@qq.com

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

gm19900510 commented 4 years ago

请更新代码，模型进行替换

------------------ 原始邮件 ------------------ 发件人: "崔勇"<notifications@github.com>; 发送时间: 2020年1月16日(星期四) 下午4:50 收件人: "Seanlinx/mtcnn"<mtcnn@noreply.github.com>; 抄送: "1025304567"<1025304567@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [Seanlinx/mtcnn] facial landmark regression (#37)

hi,i also need the code of facial landmark detection.

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"gm19900510"<notifications@github.com&gt;; 发送时间:&nbsp;2020年1月2日(星期四) 下午3:28 收件人:&nbsp;"Seanlinx/mtcnn"<mtcnn@noreply.github.com&gt;; 抄送:&nbsp;"Richar"<359894979@qq.com&gt;;"Comment"<comment@noreply.github.com&gt;; 主题:&nbsp;Re: [Seanlinx/mtcnn] facial landmark regression (#37)

hi Rainy000, could you please send me a copy of facial landmark detection? My email:1025304567@qq.com

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.