search

guoqiangqi / PFLD

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

Some problems about concat feature #44

Open douyh opened 5 years ago

douyh commented 5 years ago

The author concat last 3 feature maps and then use a FC layer to get the landmarks.

In your code model2.py, gloabal average pooling is added to the last 3 feature maps. And then you concat them.

I did the same things as you. My last 3 feature maps are 8*8 4*4 2*2. I used GAP and Concat. But the performance is not as good as what I get only using the last feature map 2*2(GAP+FC).

I use smooth L1 loss without any weights. Have you tried to use the last feature map instead of concat?

Thanks for your reply.

guoqiangqi commented 5 years ago

I did not use the last feature map instead of concat ,so i don`t konw which one is better.But in my opinion,the concat operation should help sending more information to the input of fc layer,it would improve the difficulty of model training at the same time, the poor performance may cased by the unconvergence of the model,i guessed it.