open-mmlab/mmdetection ? reverse-engineered prompt

Build me an open source computer vision toolkit for object detection in Python, with the kind of flexible setup where I can plug together different model parts, train on my own dataset, and run evaluation and demos easily.

It should support the usual detection workflows like object detection, instance segmentation, panoptic segmentation, and semi supervised training, and include a few strong baseline models so people can start from a config and fine tune instead of building everything from scratch. I’d like it to feel polished, with clear docs, example configs, sample commands for training and testing, and a simple demo for running predictions on images or videos.

If you need to check current best practices or upstream docs while wiring it up, do that. Focus on making it practical and easy to extend, with good defaults and a clean model zoo style structure.