New C++ Engine Delivers Real-Time 3D Full-Body Motion Capture From a Single Camera With Zero Python Dependencies
Summary
A new standalone C++ engine called SAM3DBody-cpp now delivers real-time 3D full-body motion capture from a single camera with zero Python dependencies, supporting multi-person tracking, a 70-joint skeleton, and direct BVH export compatible with Blender.
Key Points
- SAM3DBody-cpp is a standalone C++ inference engine that performs real-time 3D full-body reconstruction from a single camera using ONNX Runtime and ggml, requiring zero Python dependencies at runtime and supporting multi-person detection with a 70-joint skeleton including hands.
- The pipeline runs YOLO person detection, a DINOv3-ViT-H backbone, and a transformer decoder to directly regress 3D body model parameters from image features, outputting full mesh vertices, 3D keypoints, and BVH motion-capture files compatible with Blender and other DCC tools.
- An offline multi-pass processing mode is available for video files, offering zero-phase Butterworth smoothing, scene-cut detection, gap interpolation, jitter correction, and stable multi-person identity tracking to produce higher-quality BVH exports compared to the real-time online pipeline.