PN-Net

reproduce paper PN-Net:Conjoined Triplet Deep Network for Learning Local Image Descriptor

有2点需要声明

There are two important detail needed to be clarified

1 In author raw Lua code work ,I couldn't explictly find there is a batchnormlization.Maybe I dont konw how to use Lua and Torch

1 作者开源的Lua源码中没有显示的发现作者使用的BN,但是在复现中使用了BN

The reason to use BN

Because in the procedure of optimization,there is always a gradient explode and thus cant continue to training,I dont kown why .I provide two solutions:First,use gradient clip in the train code which i comment it as i have used BN,once use BN,there is no optimization problem.Second,use BN

I only do training on Brown Dataset Liberty

testing on the other two subsets Notredame and Yosemite

results

Notredame ROC

Yosemite ROC

result

We find there are 2.5 points higer than raw paper.Because we use 50w pairs,however original paper using 120w pairs.