CharlesShang / TFFRCNN

FastER RCNN built on tensorflow
MIT License
874 stars 418 forks source link

Problem with OHEM modification #64

Open hiz0ey opened 7 years ago

hiz0ey commented 7 years ago

Hi, Charles, I'm going through your code for OHEM. And I found that the only difference between OHEM version and Standard FRCNN is on the build_loss() function, right? And all these modifications are implemented on sampled rois, which makes me quite confused. According to the paper and OHEM impletation on caffe, the hard examples should be selected from all rois.