matterport/Mask_RCNN ? reverse-engineered prompt

Build me a simple Mask R CNN app that can detect objects and draw both the bounding boxes and the pixel level masks on an image. I want to be able to drop in my own photo, run a pre trained model, and see the results clearly overlaid on the picture, with labels and confidence scores.

Also make it easy to train on a custom dataset, starting from either the COCO weights or ImageNet weights, and include a small example dataset so I can test the full flow without doing a lot of setup. If it helps, add a few notebooks or an easy demo page that lets me inspect the detection steps, like the proposals, final boxes, and masks.

Keep it well organized, documented, and easy to extend. Use the current best practices for the underlying Python deep learning stack, and look up current docs online if you need to.