Xuan Lin
Xuan Lin
@quickj Here's some discussions about training landmarks #25
The speed is about 50ms per image for VGA images(30ms-40ms per image for fddb images) on titan x.
fcn @KeyKy
Yes. And it'll be faster if you set MXNET_ENGINE_TYPE to NaiveEngine
Use code on FDDB website
The recall of p-net and r-net are slightly lower(1%～%2, minimum detect face size=24) than the original model while having the same false positives, so I use lower threshold to keep...
I have updated the final disROC. The performance will be better if joint landmark detection in training according to the paper.
@mychina75 it's for testing. The thresholds I used for data preparation are a little lower just to cover more positive samples.
@hust-kevin They are the indices of valid training examples in a batch. cls_keep_inds: hard examples of positives and negatives bbox_keep_inds: positives and part faces
The input of pnet is square already