How to know if the faster R-CNN is trained to a good state according to Mini-Batch loss and accuracy trends

Hello, I am trying to train a faster R-CNN detector using my dataset, but the result is much worse than yolo V2, which is different from some papers. I think there is something wrong when training faster R-CNN. So I wonder how to know if I should change hyperparameters and when to stop training according to Mini-Batch loss and accuracy trends?

Answer 1

[Divya Gaddipati]

Hi,

You should refer to this example in trainFasterRCNNObjectDetector documentation page to understand the role of various inputs going into trainingOptions function. For one, if you are training your network using adam solver, consider using the other training options associated with it like GradientDecayFactor and SquaredGradientDecayFactor.

You could also try couple of other things:

• Make sure you have enough data for training the network.
• Try using a different base network. Refer to this link: https://www.mathworks.com/help/vision/ref/trainfasterrcnnobjectdetector.html#bvkk009-1-network
• Make sure the PositiveOverlapRange and NegativeOverlapRange are adjusted according to your data

You can also refer to the answer posted for a similar question to understand how faster R-CNN works:

https://www.mathworks.com/matlabcentral/answers/332757-faster-rcnn-code-in-matlab?s_tid=srchtitle