image.png
出自Face++和UIUC，ACMMM的。
使用了IoU Loss来替代L1 Loss做localization（regression）
之前做的笔记：
87E9CF297F71ADD297D22C8244211F4D.png
loss weight参数在文章中没有写，问了作者回答说：
In face, loss weight is: 0.001 for classification and 1 for localization
也有网友用tensorflow重现，不过用的参数不太一样：
https://github.com/zhimingluo/UnitBox_TF
这份代码训练时可能出现问题，我在issues中也有贴出过：
https://github.com/zhimingluo/UnitBox_TF/issues/2
I trained this model with the provides code, for almost 15 hours, on one NVidia 1080Ti. It consumes about 9G video memory. The final output loss from this trained network is:
Epoch [19] average Loss [12880] : 0.305612
For people that wants to train:
- prepare vgg16_weights.npz
- prepare WIDER face training images
- modify test.py's image name
网友评论