Closed mocialov closed 7 years ago
I wonder if this comment (from https://github.com/yhenon/keras-frcnn/issues/75) is relevant...:
The use of hard negative mining on bg regions is broken currently unfortunately. The issue is that the code tries to maintain a balance of negative and positive samples for training the object classifier (see lines 183-217) of train_frcnn.py. So if it gets no positive regions from the RPN it will just skip training the classifier. One option would be to add a parameter to ignore this balancing - then it could check for proposals from the RPN that overlap the bg regions and train on these?
I have generated the negative samples from the bigger test images and will train on these now
It is not an issue
Submitted an issue to the original repo: https://github.com/yhenon/keras-frcnn/issues/144