MoGe

MoGe was introduced by Microsoft Research (with USTC, Harvard, and Tsinghua collaborators) in the October 2024 arXiv paper 2410.19115 and presented as a CVPR 2025 oral. It recovers 3D geometry from a single open-domain image by directly predicting an affine-invariant 3D point map, sidestepping the ambiguous supervision that scale-dependent representations suffer from, and pairs this with a mask for regions of undefined geometry such as sky.

A ViT encoder plus convolutional decoder predicts the point map, from which depth maps and camera focal length can be derived; the model runs in about 0.2 seconds per image on an A100-class GPU and was reported to outperform prior methods across point-map, depth, and field-of-view estimation. Code and weights are released under the MIT license, and a lightweight MoGe-2 variant targets on-device inference.