openpose

Released in 2018 alongside its CVPR paper, OpenPose established the benchmark for human pose estimation performance and made the capability accessible to researchers and developers outside specialized robotics and computer vision labs.

It detects 25 body keypoints per person, covering the full skeleton from head to feet.

The system uses a Part Affinity Fields approach that detects body part locations and their connections in a single bottom-up pass across the image, then assembles complete skeleton estimates for each person.

This design enables efficient processing of scenes with many people without the computational cost scaling linearly with person counta key advantage for crowd analysis, sports broadcasting, and public space monitoring applications.

OpenPose provides a C++ library with Python bindings and runs with GPU acceleration through CUDA.

Researchers studying human motion and behavior, sports analytics teams analyzing athlete biomechanics, dance and performance artists using body tracking for interactive installations, physical therapy applications measuring range of motion, and developers building gesture recognition and activity classification systems on top of pose data use OpenPose as their foundational human pose estimation layer.

While more recent models like ViTPose and RTMPose have surpassed it in accuracy benchmarks, OpenPose's established documentation, broad platform support, and extensive community examples maintain its relevance as a reference implementation and practical tool.