SAM3DBody-cpp

Why we included this project

Most single-camera pose estimators stop at 2D keypoints, but this one goes further: it regresses full-body pose parameters straight from image features, with no depth sensor, floor plane, or stereo rig in the loop, then runs linear blend skinning to output a 3D mesh and a 70-joint skeleton that includes hands. The multi-person BVH export is what makes it interesting for animation work. It writes a standard motion-capture file per detected person, keeps identities stable across frames with an IoU tracker, and resizes each file's joint offsets to the actor's measured bone lengths, so the result drops straight into Blender or other DCC tools. The runtime is pure C++ on ONNX Runtime and ggml, so production systems can embed it without a Python dependency, though the bundled Python frontends and CSV keypoint exporter make prototyping easy. Real-time throughput assumes a CUDA GPU; on CPU-only machines it is better suited to single images or low-frequency processing.