Train Faster RCNN on Custom Dataset

Question

I am trying to run Py-Faster-RCNN for 2 class problem say, obj1 and obj2.

1. I have stored the data in Pascal format.
2. I also changed the number of output in last layers of the caffe definition to 3 ( i.e. 2 + BG ) and 12 ( 4 * 3 ).
3. For some reasons, I cannot use pretrained model.

Now, I wish to train the caffe model from scratch for above mentioned two classes. So, my question is; If I am running the following command,

./tools/train_net.py --gpu 0 --imdb voc_2007_trainval --cfg experiments/cfgs/config.yml --solver models/pascal_voc/ZF/faster_rcnn_end2end/solver.prototxt --iter 10000

Is it working as per my requirement (or, what I am thinking) i.e. object classification and localization for obj1 and obj2. Or, something else is happening? Because, my AP values are very low (obj1 = 0.042; obj2 = 0.006) for the model I obtained by running it for 10000 iterations.

Answer 1

If you have arranged your data for obj1 and obj2 classes in PASCAL format then you are correctly running the training code.

I assume that you have taken special care in creating the text files which mention about the training and testing samples. Please, see PASCAL setup for reference.