DirtyMoCap recovers 3D motion from noisy unordered markers, with 100x faster CUDA solver
jamestagg · x · 2026-10-12
A Westlake University-led team introduces DirtyMoCap (SIGGRAPH Asia 2026), a marker-layout-free motion capture framework that recovers 3D human motion from sparse, noisy, unordered marker clouds. It maps unordered observations to fixed 'proxy anchors' (joints + surface points), tracks them with a recurrent sliding-window architecture, then fits SMPL-H via a custom differentiable Gauss-Newton solver with learned confidence/smoothness/prior weights. A single model outperforms configuration-specific SOTA baselines on joints and vertices, and the custom CUDA solver gives up to 100x speedup over PyTorch. They also release HKMALA-Motion: 134 Kung Fu sequences, 3 hours, 22 styles. Paper and code are public.
More from Embodied
- Running Two AI Agents on Raspberry Pis with Full Sudo, RTL-SDRs and Starlink — natesiggard · 2026-10-12
- Unitree open-sources 6B humanoid model UnifoLM-WLA-1.0, claims fine-tuning on a 24GB GPU — lmoroney · 2026-10-12
- Tesla Cybercab tracker logs 145 on the road, 319 registered with Texas DMV — JOBhakdi · 2026-10-12
- Liquid AI open-sources Open d1 decision models: d1-3B runs at 35ms per image on Jetson Thor — TheTuringPost · 2026-10-12
- Fleet Week traffic trapped Andrew Chen's Waymo for 15 min until humans-in-the-loop rerouted it — andrewchen · 2026-10-12
- COMODO distills video intelligence into IMU sensors for camera-free egocentric AI glasses — flosalim · 2026-10-12