Can only predict one class

[RJVisee44]

Hi guys, I've trained Fast and Faster RCNN so I know a bit about the process. I believe I've followed all the steps to train Fast RCNN with my own object proposals. I can get it working for one class problems (Background, other class) but I cant get it working for two classes (background, Class1, class2).

When i test the model by running the demo.py with my object proposals and my trained Fast RCNN model, Class 1 is correctly detected, but Class 2 is detected in random areas (usually bottom left corner) where no object proposals even exist.

Is this most likely a problem during training? Do I have to specify which class the initial object proposals for training belong to or does the algorithm match them based on overlap? Also, do these need to be included in factory.py. I'd say no based on the fact that it worked for the one class problem.

It seems to correctly load the train.txt and train.mat file. I also changed the annotations to reflect the added class.

I've gone through the code a bunch and am completely lost. Any tips would help.