// the find
CMU-Perceptual-Computing-Lab/openpose
OpenPose: Real-time multi-person keypoint detection library for body, face, hands, and foot estimation
OpenPose is a CMU research library for real-time human pose estimation — body, hands, face, and feet — from images or video. It was the first system to do all of this jointly in real time, and the Part Affinity Fields approach it introduced remains widely cited. It's for researchers and practitioners who need keypoint detection and don't want to train their own models from scratch.
The constant-time body detection regardless of person count is a genuine architectural win — most competitors scale linearly with crowd size. The 135-keypoint whole-body coverage (body + 2 hands + face + feet) in a single pass is still hard to beat without switching to newer transformer-based approaches. Both C++ and Python APIs are available, and the tutorial examples are numerous and well-organized, covering async pipelines, multi-GPU, custom input/output, and calibration. The documentation is unusually thorough for a research repo — installation guides for multiple OS/hardware combinations, UML diagrams of the threading model, and a proper FAQ.
It's built on Caffe, which is effectively unmaintained and increasingly painful to build in 2024 — CUDA compatibility is a recurring headache and you'll spend real time fighting the build before running a single inference. Last meaningful commit was 2024 but the underlying model architecture is from 2017–2019; newer approaches like ViTPose or RTMPose are meaningfully more accurate, especially on occluded or unusual poses. The license is non-commercial only, which rules it out for most production use cases without a CMU commercial agreement. Hand and face detection runtime scales with person count even though body detection doesn't — the README buries this inconsistency.