ProxyPose lands at NeurIPS: track 6-DoF pose from a single query pixel

CSProfKGD · x · 2026-09-26

Researchers from the University of Toronto and Vector Institute present ProxyPose, accepted to NeurIPS 2026. Given a monocular RGB video and one query pixel, a LoRA-finetuned video generative model first translates the input into a proxy video where a colored cube mimics the point's local rigid-body motion; classical learning-free geometry (contour fitting + Perspective-n-Point) then recovers the full 6-DoF pose trajectory. The approach extends to camera pose tracking, with paper, code, and webpage available.

Original post →

More from Embodied

Embodied channel →