I followed the steps of a tutorial for instance segmentation using Detectron 2 with my own dataset (the annotations were created with ImageJ but were succesfully converted to COCO format with a python script). I have 108 training images with only 1 class (multiple instances of this class in each image).