Post by Sensing, Interaction & Perception Lab · ETH Zürich

Wearable motion capture is prone to inertial drift. Our #CVPR2026 paper turns between-sensor distances into geometric guidance for diffusion-based pose estimation for more accurate and smooth motion and much reduced joint-positioin error. Current sparse inertial mocap systems are practical and camera-free, but they struggle with inaccurate limb placement due to IMU drift. Our method 𝗨𝗹𝘁𝗿𝗮 𝗗𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗣𝗼𝘀𝗲𝗿 leverages wearable IMU+UWB trackers for inertial sensing and modeling the geometry behind between-sensor distances. We first reconstruct the 3D layout of the body-worn sensors and then use it to guide a diffusion model toward physically consistent full-body poses. Ultra Diffusion Poser 👍 improves pose accuracy: reducing joint position error by up to 22% over previous SOTA. 📡 uses only wearable sensors: 6 body-worn IMU+UWB nodes—no cameras, markers, or external infrastructure. 🧍 improves limb placement: UWB distances provide strong on-body constraints for wrists, knees, head, and pelvis. ⚡ runs in real time: neural inference avoids expensive post-hoc physics optimization and can run substantially faster than prior optimization-based pipelines. Ultra Diffusion Poser explicitly reconstructs the 3D arrangement of the body-worn sensors from the measured UWB distances. This turns noisy pairwise ranging measurements into a geometric representation of how the sensors are positioned relative to each other. We combine this geometric information with inertial cues and motion history to estimate full-body motion with a diffusion model. During inference, UDP uses distances to continuously guide pose generation, encouraging predictions that remain consistent with the sensor measurements and improving both accuracy and physical plausibility. Dominik Hollidt, Tommaso Bendinelli, and Christian Holz. Ultra Diffusion Poser: Diffusion-Based Human Motion Tracking From Sparse Inertial Sensors and Ranging-Based Between-Sensor Distances. Conference on Computer Vision and Pattern Recognition (CVPR 2026). arXiv: https://lnkd.in/eW3X-jQ8 code: https://lnkd.in/eTER_zHK full 7-min video: https://lnkd.in/enHBJybq project page: https://lnkd.in/eFpH6FBa Department of Computer Science (D-INFK), ETH Zürich • ETH Zürich #MotionCapture #WearableTechnology #HumanPoseEstimation #DigitalHumans #UWB #IMU #DiffusionModels #XR #HumanComputerInteraction

Video Content