I have a dataset of high-resolution drone images, which I have split into tiles of size 512x512. The tiles will be annotated by delineating a specific type of vegetation and stored in COCO format. I would like to use self-supervised learning for instance segmentation in the tiles. The Pytorch Lightning bolts module provides already pre-trained self-supervised models, such as SimCLR. According to the bolts module documentation, it should be possible to finetune SimCLR on any computer vision task, such as instance segmentation, using bolts.
The only examples I found, however, show how SimCLR is finetuned on classification tasks. How would one finetune SimCLR (or another self-supervised model) on an instance segmentation task?