Jie Zhang

@hesperxxz your formula is right. And I think I'm using number_of_positive_samples = 100,000 (not all of them has landmark points). And I only trained 2700 carts. The FPR you calculates...

Hi @oylz , My understanding of down sampling is scale the image area. So I choose 3 different scales by image's size not the width or height.

Hard negative mining is to find the false positive samples in background images. You are short of background images too badly. The reset times shouldn't be too large, usually, reset...

The problem here is that hard negative mining process can not generate enough negative patches from background images. This means your model trained at this time is good enough to...

@ardeal Sorry, I don't have a trained model. I may also change the details of the source code. However, the code at this time should work well.

You should prepare much more data than the paper described. You also need a very powerful machine with many cpu cores to run the algorithm. Since my lack of these...

They are described in README

1. The threshold in Regression is random. If you iterate all threshold [-255, 255], it will cost too much time, while random threshold can still give you a pretty good...

这个最好是你的数据集本身就包含有 landmark 数据，不然的话，你可以找一些工具或者开放的 sdk 等自己给数据标上 landmark。 http://ibug.doc.ic.ac.uk/resources/300-W/ http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

@tonyzzzzz 点必须是全的，不能是部分点