search

guoqiangqi / PFLD

Implementation of PFLD A Practical Facial Landmark Detector , reference to https://arxiv.org/pdf/1902.10859.pdf
627 stars 166 forks source link

Are Euler Angle GT and W_n pre-defined? #12

Closed PlumedSerpent closed 5 years ago

PlumedSerpent commented 5 years ago

According to my understanding, it seems we can calculate the Euler Angle GT and w_n of all samples, and return them in dataset iterators. Right?

guoqiangqi commented 5 years ago

The euler angles can be calculated in the preprocess procedure ,but w_n is not suitable ,it is different in each batch,it should be kepping updating during trianing.

PlumedSerpent commented 5 years ago

The euler angles can be calculated in the preprocess procedure ,but w_n is not suitable ,it is different in each batch,it should be kepping updating during trianing.

I re-implement PFLD in pytorch, and I do the pre-processing procedure just as you said. But I found a confusing result, it seems the Euler predict branch doesn't contribute to the accuracy. In detail, I got 5.7% NME w/o the Euler prediction and weight_attributes loss in WFLW dataset. (btw, I trained 2k epochs, which may be too much in my thought.) In your experiments, did you observe the improvement brought by the weight term in loss function?

guoqiangqi commented 5 years ago

The euler angles can be calculated in the preprocess procedure ,but w_n is not suitable ,it is different in each batch,it should be kepping updating during trianing.

I re-implement PFLD in pytorch, and I do the pre-processing procedure just as you said. But I found a confusing result, it seems the Euler predict branch doesn't contribute to the accuracy. In detail, I got 5.7% NME w/o the Euler prediction and weight_attributes loss in WFLW dataset. (btw, I trained 2k epochs, which may be too much in my thought.) In your experiments, did you observe the improvement brought by the weight term in loss function?

I did not compare the results with ones without euler angles prediction , but if you have taken the euler angles prediction loss as a part of total losses,it should have influence on the common part of whole model .It is clean that multi-loss is also efficient to improve the model performence.

PlumedSerpent commented 5 years ago

The euler angles can be calculated in the preprocess procedure ,but w_n is not suitable ,it is different in each batch,it should be kepping updating during trianing. I re-implement PFLD in pytorch, and I do the pre-processing procedure just as you said. But I found a confusing result, it seems the Euler predict branch doesn't contribute to the accuracy. In detail, I got 5.7% NME w/o the Euler prediction and weight_attributes loss in WFLW dataset. (btw, I trained 2k epochs, which may be too much in my thought.) In your experiments, did you observe the improvement brought by the weight term in loss function?

I did not compare the results with ones without euler angles prediction , but if you have taken the euler angles prediction loss as a part of total losses,it should have influence on the common part of whole model .It is clean that multi-loss is also efficient to improve the model performence.

OK, thanks for your reply, I'll try more experiments. Because I have been clear about the data pre-processing problem, I just close this issue.

DeepJake commented 5 years ago

The euler angles can be calculated in the preprocess procedure ,but w_n is not suitable ,it is different in each batch,it should be kepping updating during trianing.

I re-implement PFLD in pytorch, and I do the pre-processing procedure just as you said. But I found a confusing result, it seems the Euler predict branch doesn't contribute to the accuracy. In detail, I got 5.7% NME w/o the Euler prediction and weight_attributes loss in WFLW dataset. (btw, I trained 2k epochs, which may be too much in my thought.) In your experiments, did you observe the improvement brought by the weight term in loss function?

I try to re-implement PFLD in pytorch but the accuracy is worse than this tf-version. What about your model? Sorry to bother you but may I share your pytorch code?

xubin19939 commented 5 years ago

@guoqiangqi 看您的代码里eular angle 不会直接读出的gd的[203:206]三位吗，这三个值不就是eular 角度？为什么上面还提到了额外的计算？

guoqiangqi commented 5 years ago

这是更新后的代码，在图像读入时计算的euler angles。

425418567@qq.com

From: xubin19939 Date: 2019-08-15 11:06 To: guoqiangqi/PFLD CC: Guoqiang QI; Mention Subject: Re: [guoqiangqi/PFLD] Are Euler Angle GT and W_n pre-defined? (#12) @guoqiangqi 看您的代码里eular angle 不会直接读出的gd的[203:206]三位吗，这三个值不就是eular 角度？为什么上面还提到了额外的计算？ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.