Harvard's XM4D, RSS 2025 Best Systems Paper, does certifiable 4D reconstruction from monocular video
kwangmoo_yi · x · 2026-11-07
XM4D, from Harvard with UC Berkeley (conference version XM won RSS 2025 Best Systems Paper Award), splits 4D reconstruction into certified stages:
- Input: a casual monocular video; off-the-shelf point trackers give 2D tracks, a monocular depth model gives per-frame depth, masks split tracks into static/dynamic.
- Solve: a GPU solver first certifies a scaled bundle adjustment over static tracks as globally optimal, then certifies the deformation graph's motion—provable optimality rather than plain iterative optimization.
- Render: certified cameras and motion initialize dynamic 4D Gaussian splatting, refined for novel-view rendering.
The philosophy: certified geometry first, rendering second. Paper and code are open.
More from Embodied
- ROS weekly: ROSCon Italy 2026 a hit, ManyForge and Lopsy projects debut — kscottz · 2026-11-07
- Dream-1 update adds SO-101 and Franka support; user runs it zero-shot on a Roomba — scott_e_reed · 2026-11-07
- Eka Robotics heads to CoRL and opens hiring for its technical team — pulkitology · 2026-11-07
- XM4D: certifiable geometric optimization for 4D reconstruction from monocular video — kwangmoo_yi · 2026-11-07
- NVIDIA's Lyra 2.0 turns a single image into explorable, consistent 3D worlds — NVIDIAAI · 2026-11-07
- Cartwheel's Swing Choreo model makes robots perform any pose from a text prompt — andrew_n_carr · 2026-11-07