I want to use UPSNet (github repo) (paper) to train a model to perform panoptic segmentation on my own dataset. I would also consider using a model based on Mask-RCNN to simply perform instance segmentation.
These models are typically ran on standard datasets like Cityscapes and the COCO dataset, and I can find very little about using your own data online.
I hope to assign for each pixel either a background or target class label, and if the pixel is in the target class, an integer instance ID number. My own data, at this point, consists of $N_{images}$ each of size (512, 512). I have these collected into a NumPy array of shape ($N_{images}$, 512, 512, 2), where channel 0 of the last dimension represents a ground truth background/target mask, and channel 1 stores the integer id (ranging from 1 to $N_{objects}$). How do I transform this in order to obtain a dataset similar enough to COCO that I could try to run predictions using the prebuilt models on it?
