paddlepaddle/paddledetection

PaddleDetection is a Python toolkit for building computer vision systems that can locate and identify objects in images and video. It is developed by Baidu and built on top of PaddlePaddle, Baidu's deep learning framework. The toolkit covers four main tasks: object detection (drawing boxes around things like cars, people, or products), instance segmentation (coloring in the exact pixel outline of each object rather than just boxing it), tracking objects as they move across video frames, and detecting body keypoints such as elbows and knees for human pose analysis. The project includes a large library of pre-trained models covering a wide range of published research approaches, from lightweight models fast enough to run on mobile devices to high-accuracy models for server-side processing. Several of these are branded as PP-models, meaning Baidu's own optimized versions: PP-YOLOE for general detection, PP-PicoDet for very lightweight real-time detection, PP-Tracking for multi-object tracking, PP-TinyPose for body keypoint detection, PP-Human for pedestrian analysis (counting people, detecting intrusions, recognizing activities), and PP-Vehicle for vehicle analysis (license plate recognition, traffic counting, violation detection). The toolkit is designed to take a developer from raw data through to a deployed model in one place. It handles data preparation, model selection and training, and multiple deployment options including standard server deployment, serving the model as an API, and deploying to edge devices. It supports switching between different hardware targets including Nvidia GPUs and various Chinese AI chips. The README and most documentation are written primarily in Chinese, with an English version of the README available as a separate file. The project is licensed under the Apache 2.0 license and targets Python 3.7 or newer, running on Linux, Windows, and Mac. The full README is longer than what was shown.