PaddlePaddle/PaddleDetection ? reverse-engineered prompt

Build me a simple but solid computer vision toolkit using PaddlePaddle that can detect objects in images and video, and also handle instance segmentation, multi object tracking, and pose keypoints.

I want it to be easy to try on my own photos or webcam, with a clear way to pick a model, run inference, and save the results. Please include a couple of ready to use demos for common cases like general object detection, person detection, and tracking, plus basic training and evaluation support so I can swap in my own dataset later.

Make the project feel polished, with good defaults, clear README instructions, and sensible folder organization for configs, demos, deployment, and tests. If you need to check current PaddlePaddle or model usage details, look them up online first. Keep it lightweight enough that I can get something working quickly, but complete enough that it feels like a real production friendly starting point.