Hello,
One way to improve the accuracy, in my opinion, is to train, say, Unet neural network (see how to do that here ) on some data, for example, public dataset LPIS 2016 for semantic segmentation and then use WatershedComponents similar what you did to go to instance segmentation.