Uni3R unifies 3D reconstruction and semantic understanding from unposed views

rsasaki0109 · x · 2026-07-25

Uni3R unifies 3D reconstruction and semantic understanding

This CVPR 2026 highlight presents Uni3R, a method for generalizable Gaussian splatting from unposed multi-view images.

• It aims to recover 3D geometry, novel-view rendering, depth, and segmentation in one pipeline.
• The visual shows reconstruction on indoor scenes and outputs for RGB, depth, and semantic segmentation.
• The paper also reports benchmark results on datasets such as ScanNet, RE10K, and ACID.
• The method is shown to run in about 0.15 seconds in the illustrated example.

Original post →

More from Embodied

• Qualcomm-powered robot dies mid-demo and gets the perfect “I quit” punchline — tristanbob · 2026-07-25
• Tesla says trailing-12-month revenue topped $100B as Cybercab production starts — mitchdeg · 2026-07-25
• Plant 3D scan demo reaches 8 million splats with 1,080 photos — janusch_patas · 2026-07-25
• Ecovacs open-sources its Bajie embodied robot platform with hardware, source code, and local 7B model support — 量子位 · 2026-07-25
• Sunday Robotics says its home robot beta build is now safer and more capable — tonyzzhao · 2026-07-25
• Robotics policy keeps the vision backbone frozen and trains on a consumer GPU — mayfer · 2026-07-25

Embodied channel →