--- title: ExMR - Extremity Motion Refinement emoji: 🕺 colorFrom: indigo colorTo: gray sdk: gradio sdk_version: 6.30.0 app_file: app.py short_description: 3D body, hand and foot motion from monocular video python_version: "3.12" startup_duration_timeout: 1h --- # ExMR: Extremity Motion Refinement for Monocular Human Motion Capture Reconstruct 3D body, hand and foot motion (SMPL-X) of the main person in a monocular video with [ExMR](https://huggingface.co/papers/2610.26213) ([code](https://github.com/twehrbein/ExMR)). Upload a short clip (≤ 30 s, one dominant person) — the app tracks the person, extracts whole-body 2D keypoints with Sapiens2-1B, body tokens with SAM 3D Body (DINOv3-H+) and per-hand tokens with HaMeR, runs the ExMR transformer (with its Foot and Hand Refiners), and renders the input-view and world-view meshes side by side, plus a `results.npz` with the raw SMPL-X parameters. Deviations from the official pipeline (ZeroGPU constraints): SimpleVO visual odometry runs synchronously in a background *thread* instead of a child process (the official code uses a `ProcessPoolExecutor`, which is not allowed inside a ZeroGPU worker), and inputs are capped at 900 frames. The static-camera path (the default) skips visual odometry entirely, matching the authors' example. ## Licenses The ExMR code (`lib/exmr/`), SAM 3D Body, Sapiens2, the SMPL-X and MANO body models, and HaMeR ship under their own research-use licenses — see `licenses/`. The example video is from [Pexels](https://www.pexels.com/video/woman-doing-dancing-exercise-in-the-garden-3690649/) (Pexels license, free to use).