facebookresearch/detectron2 ? reverse-engineered prompt

I want to build a powerful computer vision library in Python, something like Facebook's Detectron2. The main goal is to run object detection and segmentation on images. It should be able to do basic stuff like drawing bounding boxes around objects, but also more advanced things like creating a pixel-perfect mask outlining each object's shape, and even detecting human poses and keypoints.

This should be a proper library, not just a script. It needs to have a "model zoo" with a bunch of state-of-the-art, pre-trained models that I can easily download and use. I also need the ability to train these models on my own custom datasets. Please build it on top of PyTorch, and make sure it's fast and well-structured for research and building other projects. Also include a way to export the finished models for deployment.