Instance segmentation on custom dataset with COCO annotations

I would like to train an instance segmentation model on a custom dataset, for which I converted the annotations in COCO format. Which frameworks would you suggest me to use? I know about torchvision, MMDetection and Detectron2. I need a framework which

  1. supports instance segmentation (clearly :slightly_smiling_face:)
  2. consumes COCO annotations
  3. is a good compromise between flexibility and robustness/ease of use.
  4. has a Docker image. Bonus points if the Docker image is available on, or if it’s compatible with the NVIDIA PyTorch image available there.

Which one would you suggest to use? Thanks,