search

reshow / PRNet-PyTorch

The training and evaluation code for PRNet (《Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network》)
73 stars 16 forks source link

get a bigger NME result #18

Open Nancyh-e opened 3 years ago

Nancyh-e commented 3 years ago

sorry to bother you and thanks for release your version on github. But i have some questions about NME. when i try to get the NME of PRNET, i got a really bigger result than the author. so i doubt i use a wrong method. How could I get the NME result of 3D 68 points on AFLW 2000 data? I try to calculating the L2 distance between the predicted result of 3D model 68 points(from kpt.txt)and ground truth(from AFLW2000 .mat file of pt3d_68) , and then use the result to /d(d is the bounding box of 68 kpts, it is about 180)/68(the number of points) according to NME expression.The method to get d is llen = sqrt((maxx - minx) * (maxy - miny)); or d = np.sqrt((maxx - minx) 2 + (maxy - miny) 2).(don't know choose which one) But the 2 methods to get d is incorrect after i get the final result of NME.I doubt it is a mistake,and I don't know how to calculate.The result is similar to 0.33 or so ,not similar to the paper result like 0.0326. Many thanks for taking your time to read it, i am really looking forward to your reply! ! !Have a great day! Thanks again!:）and my mailbox is CYNh_e@126.com.

reshow commented 3 years ago

Sorry to reply so late. I use the first one to calculate NME. The y-axis direction of the processed position map is not the same as the initial .mat file. (I used a y=256-y in the data pre-process.)

I suggest you to plot the landmarks on the images to check it.