I validated the model on aflw2000-3d database, get bigger NME.

yfeng95 / PRNet

Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network (ECCV 2018)

http://openaccess.thecvf.com/content_ECCV_2018/papers/Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf

MIT License

4.96k stars 945 forks source link

I validated the model on aflw2000-3d database, get bigger NME. #182

Open lijie2160 opened 4 years ago

lijie2160 commented 4 years ago

the result is: [ 0, 30] Mean: 7.839, Std: 2.902 [30, 60] Mean: 9.314, Std: 4.167 [60, 90] Mean: 13.345, Std: 8.840 [ 0, 90] Mean: 10.166, Std: 2.327

is bigger than the paper mentioned.

Shiyu555 commented 4 years ago

hi , could you tell me how to calculate 3D NME?I don't know when I calculate 3D NME,how can I get bbox. bbox_size = torch.sqrt((right - left) * (bottom - top)) is right?

Nancyh-e commented 3 years ago

hi , could you tell me how to calculate 3D NME?I don't know when I calculate 3D NME,how can I get bbox. bbox_size = torch.sqrt((right - left) * (bottom - top)) is right?

i have the same question withyou, how to calculate NME?that makes me big head hhhh:)

lijie2160 commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

Nancyh-e commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

ok! I'll try it! thanks a lot!Have a great day!!! :)

Nancyh-e commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

hi,I am sorry to bother you again,do you get a similar NME result( according to the link method you have sent to me) to the author's PRNet paper? I use the predicted result of 3D model 68 points(from _kpt.txt of some AFLW2000 test images )and ground truth(from AFLW2000 .mat file of pt3d_68), are they correct data to calculate the NME? Really hope to receive your reply. thanks very much! :) My mail box is CYNh_e@126.com 求大佬回复一下啊啊啊啊救救孩子

Shiyu555 commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

hi,I am sorry to bother you again,do you get a similar NME result( according to the link method you have sent to me) to the author's PRNet paper? I use the predicted result of 3D model 68 points(from _kpt.txt of some AFLW2000 test images )and ground truth(from AFLW2000 .mat file of pt3d_68), are they correct data to calculate the NME? Really hope to receive your reply. thanks very much! :) My mail box is CYNh_e@126.com 求大佬回复一下啊啊啊啊救救孩子

I wonder what is _ kpt.txt ? The third dimension of position map presents the spatial information ，I calculate NME directly with the predicted position map and the real position map.

lijie2160 commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

hi,I am sorry to bother you again,do you get a similar NME result( according to the link method you have sent to me) to the author's PRNet paper? I use the predicted result of 3D model 68 points(from _kpt.txt of some AFLW2000 test images )and ground truth(from AFLW2000 .mat file of pt3d_68), are they correct data to calculate the NME? Really hope to receive your reply. thanks very much! :) My mail box is CYNh_e@126.com 求大佬回复一下啊啊啊啊救救孩子

我只比较的前两维的数据，没有加第三维

Nancyh-e commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

hi,I am sorry to bother you again,do you get a similar NME result( according to the link method you have sent to me) to the author's PRNet paper? I use the predicted result of 3D model 68 points(from _kpt.txt of some AFLW2000 test images )and ground truth(from AFLW2000 .mat file of pt3d_68), are they correct data to calculate the NME? Really hope to receive your reply. thanks very much! :) My mail box is CYNh_e@126.com 求大佬回复一下啊啊啊啊救救孩子

我只比较的前两维的数据，没有加第三维

大佬，您是咋比较计算的？我是进行人脸重建运行demo.py，从position map获取了三维数据点存进了.txt文件，然后和AFLW2000数据集的.mat文件的三维数据进行比较，但是不是很对，前两维也不对，得不出作者的结果，可以指点一下吗？大佬方便留下联系方式吗，啊啊啊这个困扰我好久了！万分感谢！

Nancyh-e commented 3 years ago

@Shiyu555 @Nancyh-e HERE: https://github.com/cleardusk/3DDFA/blob/master/benchmark_aflw.py

hi,I am sorry to bother you again,do you get a similar NME result( according to the link method you have sent to me) to the author's PRNet paper? I use the predicted result of 3D model 68 points(from _kpt.txt of some AFLW2000 test images )and ground truth(from AFLW2000 .mat file of pt3d_68), are they correct data to calculate the NME? Really hope to receive your reply. thanks very much! :) My mail box is CYNh_e@126.com 求大佬回复一下啊啊啊啊救救孩子

I wonder what is _ kpt.txt ? The third dimension of position map presents the spatial information ，I calculate NME directly with the predicted position map and the real position map.

it is a txt file to put the 3D 68 points predict datas from position map(in demo.py can find it). and how to use the predicted position map and the real position map to calculate NME directly ? Did you get a similar result to author?thanks a lot, hope your relpy!!!

rakacendekia commented 3 years ago

@lijie2160 @Nancyh-e hii, could you please send the NMEevaluation code that you have written to me. I have tried it, but there are still errors. thank you My mail box is rcendekia007@gmail.com