Improve Mask R-CNN segmentation

I’ve been using Mask R-CNN on my project and I’m very happy with the detection results. However, I’m not fully satisfied by the segmentation as I don’t think it is precise enough for my needs.

I tried to modify the MaskRCNNPredictor without success. So:

  • What can I do to improve the segmentation part?
  • Can I change the loss?
  • Can I modify other parts of the model?

Thank you for your help