OCH3R: Object-Centric Holistic 3D Reconstruction

Published in CVPR OpenSUN3D Workshop, 2026

OCH3R is a unified framework for object-centric holistic 3D reconstruction from a single RGB image. In one forward pass, a transformer simultaneously predicts all object instances together with their 6D poses and detailed 3D reconstructions, producing per-pixel semantic embeddings, depth, object coordinates, and 3D Gaussian representations. By transforming Gaussians into canonical space with the predicted poses for supervision, it avoids expensive per-image label generation, and achieves state-of-the-art results on indoor benchmarks with fully feed-forward inference that scales independently of the number of objects.

Recommended citation: Du, Y., You, Y., Wan, X., & Guibas, L. (2026). OCH3R: Object-Centric Holistic 3D Reconstruction. CVPR 2026 OpenSUN3D Workshop.