Could you provide your training errors on each network?

[windpls]

Hi, thanks for your nice code. Recently I'm trying to implement the algorithm using Caffe. Could you provide your training errors on 12/24/48 net/calibration net each so I can check whether my code is correct?

[layumi]

@windpls 12net train-error about 5% 24net train-error about 2% 48net train-error about 1% 12netc train-error about 50% 24netc train-error about 30% 48netc train-error about 20% I am sorry. I didn't remember the number clear.

[windpls]

hi, @layumi , thanks for your reply.. But I have another two questions:

1. Did you do data augmentation for positive samples? I tried some methods like mirroring the image, jittering the bounding boxes, but all these strategies increased the validation error when training 12net, resulting in an validation accuracy about 92.7%, comparing with 99.0% without data augmentation.
2. How did you decide the thresholds of each nets? The paper suggests that the detector should get 99% recall rate on the subset of AFLW after 12net and 12 calibration net. So I leaved out 10% images of AFLW. But the way to measure the recall rate confused me. Do I need to use multi-scale detection? And is the NMS needed?

Expect your reply.

[layumi]

@windpls 1.I use mirroring the image and adding dropout 2. I test model on fddb and tune the threshold according to the recall rate in paper. after 12net 12netc 94% after 24net 24netc 89% after 48net 48netc 85%

[windpls]

@layumi I passed an image through your 12net and 12netc network and found that lots of false alarm boxes were remained. The paper said that only 2.36% detection windows passed the 12-net and 12-c net, so it seems that there is a gap between your implementation and the author's 12net, isn't it?

[layumi]

@windpls after 12net 12netc and nms it remains average about 1000(this number I don't remember clearly) after 24net 24netc and nms it remains average about 100(this number I don't remember clearly) after 48net 48netc it remains average about 30 after 0.25 nms 4 windows left.