Mira-Scene: Pixel-Aligned Layouts for Generative 3D Scene
Mira-Scene turns object placement into a dense pixel-to-3D correspondence problem instead of sparse pose regression.
The paper introduces a Canonical Coordinate Map, which maps each visible object pixel into a bounded object coordinate space. Paired with a scene-space point cloud map, those correspondences are used to recover object transformations through geometric alignment. The authors say this lets the system train from scalable object-level 3D data without scene-level layout annotations. Reported tests across indoor, outdoor, synthetic, and in-the-wild scenes show layout gains over SAM3D, including 39.8% relative improvement in 3D-IoU.
HF Daily Papers' note
The paper introduces a Canonical Coordinate Map, which maps each visible object pixel into a bounded object coordinate space. Paired with a scene-space point cloud map, those correspondences are used to recover object transformations through geometric alignment. The authors say this lets the system train from scalable object-level 3D data without scene-level layout annotations. Reported tests across indoor, outdoor, synthetic, and in-the-wild scenes show layout gains over SAM3D, including 39.8% relative improvement in 3D-IoU.
HF Daily Papers' note
score 4