Training instance segmentation of seedlings. Better results with different backgrounds?

Question

I want to train a Masked RCNN neural network to be able to segment different parts of the seedlings. We have images of the seedlings with black and white background, would it be best to train the CNN with only one background or both of them combined, to get good segmentation results? We will only acquire images with black background in the final pipeline.

Answer 1

It’s best to have your training data match your production inputs. If you are always going to have your seedlings on a black background during inferencing, you don’t need to have them on a white background during training.