How to train pretrained Faster RCNN with oriented bounding boxes

I am trying to train pretrained faster rcnn model with oriented bounding boxes. Dataset I use is DOTA. All instances are annotated by oriented bounding boxes. Is there any way in pytorch to train faster rcnn with oriented bboxes? Or can anyone suggest me a way to do so?

Thanks for any help.

1 Like

Edit: As long as I understand, there are two different approaches to detect rotated objects in images. The first one is to use rotational rpn (RRPN) which is a modified RPN network that produces inclined proposals and anchors. The second one is to use R2CNN which is a faster rcnn with an extra regression branch for the calculation of angle.

As I searched, pytorch does not have any built-in implementation for these networks. Moreover, I also couldn’t find proper and basic implementation details on the internet. So, is it easy to implement an extra regression branch to faster rcnn to predict angle? If so, could someone give me a basic understanding of how to modify the regression branch of the faster rcnn to predict angles of boxes? I assume adding extra regression branch is easier than to create RRPN.

Or to put it simply, how can add oriented bounding box regression to faster rcnn?