I have a sample that consists of two partitions: one is upright, one is upside down. The image bellow.

My downstream task is object detection. The object in upside down partition fail to be detected. Right now, train a new model with rotation augmentation is not a choice.

I'm still thinking about (1) how to automatically detect if there is any partition is not upright,

(2) split the image and (3) upright the partition before feed to the object detection model.

Any idea is really appreciated!