Question about accuracy of the model about pfld HOT 19 CLOSED

@guoqiangqi OK, I see it. You has calculated it by adding the GT of attribute classes. Thanks!

from pfld.

from pfld.

Hi，
I found this code is for training on 98-pts dataset. Do you have trained on 68-pts dataset 300-W and get the same accuracy as the PFLD paper said?

i have no trained with 300-W database ,so i did not compare the accuracy between the paper and my code ,and the rusults on WFLW(98-pts) is poor ,is sure that the LAB net should be better .

from pfld.

@guoqiangqi Thanks for your reply. I have tried on 300-W, and the result is not good either. I'll try LAB.

from pfld.

@guoqiangqi Hi, it seems that in your code you don't use auxiliary net to calculate loss. Have you tried that trick yet?

from pfld.

I have not updated the euler angles pridiction loss code ,you shou add the loss code before you train.

from pfld.

@guoqiangqi Do you use solvePnP from opencv to calculate ground-truth euler angles?

from pfld.

yeah,but my code has a flaw that i calculate euler angles ground-truth while training process,so the training speed have slowed down because some work have to be finished on the cpu ,you should calculate the euler angles in the preprocess code.

from pfld.

The related code is included in euler_angles.py

from pfld.

@guoqiangqi OK, thank you. I have noticed that. I will give a try and hope to get as good result as paper mentioned.

from pfld.

@guoqiangqi OK, thank you. I have noticed that. I will give a try and hope to get as good result as paper mentioned.
I have reviewed the code of this version ,the code for calculating euler angles prediction loss have been updated early ,you can find this part in train_model.py by searching the variable named _sum_k .

from pfld.

@guoqiangqi emmm, so the result is poor when you have added those loss terms?

from pfld.

@guoqiangqi emmm, so the result is poor when you have added those loss terms?

I evaluated the mean error and failure rate on WFLW dataset ,which has 98-pts per face ,and LAB is further better ,but PFLD is faster ,so it is hard to say which one is better.

from pfld.

@guoqiangqi Author reports both better accuracy and faster speed over LAB on 300W dataset in paper. So I expect similar result on WFLW dataset. But it seems that there is a series accuracy drop of PFLD on WFLW.

from pfld.

from pfld.

@guoqiangqi thanks for your open source work. In the paper, except for angle, the loss also has the weighting parameter WnC.

In addition, we categorize a sample into one or multiple attribute classes including profile-face, frontal-face, head-up, head-down, expression, and occlusion. The weighting parameter ω n c is adjusted according to the fraction of samples belonging to class c (this work simply adopts the reciprocal of fraction).

How to calculate it? Whether to add the GT of multiple attribute classes or calculate by the pose angle?

from pfld.

@guoqiangqi Author reports both better accuracy and faster speed over LAB on 300W dataset in paper. So I expect similar result on WFLW dataset. But it seems that there is a series accuracy drop of PFLD on WFLW.

Even that ,the results of LAB is better than PFLD on AFLW dataset ,you can compare the mean error form both paper.

from pfld.

@guoqiangqi I didn't compare the exact number from both paper and thank u for pointing out. It seems that PFLD paper has some inappropriate conclusion.

from pfld.

The WFLW dataset has considerated the complicated situation ,many faces has big angles of deflection ,and PFLD has pretrained on WFLW,so there are more images have been used for training. [email protected] From: xhq11 Date: 2019-06-03 18:04 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] Question about accuracy of the model (#1) @guoqiangqi Author reports both better accuracy and faster speed over LAB on 300W dataset in paper. So I expect similar result on WFLW dataset. But it seems that there is a series accuracy drop of PFLD on WFLW. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

Hi，did you mean the PFLD perfect result is due to the pre train on WFLW?

from pfld.