is Faster RCNN rotation invariant?

Question

I want to generate a dataset of vehicles to train in networks such as RCNN, Fast RCNN, Faster RCNN, etc. My question is:

Does the angle of vehicles in the original image matter for the detector? I mean if the dataset only contains vehicle images which are all directed from left to right, can the trained network detect any vehicle with any angle?

Answer 1

Detectron2 added Rotated Faster RCNN network recently. To create such model, you should create annotations for vehicles with rotated bounding box, which is:

rbbox = [center_x, center_x, width, height, angle]

Visit this link for more information.

Answer 2

No！ Because CNN's rotation invariance is limited, Faster RCNN which is based on CNN can't detect object whose orientaion is too different from the orientation of examples in training set.

Answer 3

The answer to this question is no. faster rcnn has only one option for augmentation which is side to side flipping. You have to add rotation to it by yourself. You can train the network with a lot of samples of rotated objects with different angles.