Researchers have developed Cross3R, a novel feed-forward model capable of reconstructing 3D point clouds and camera poses from a combination of satellite, drone, and ground images. This approach overcomes the limitations of traditional 3D reconstruction methods that rely on nadir satellite imagery, which often lack crucial roll, pitch, and altitude information. By incorporating a single UAV image as an intermediate viewpoint, Cross3R can infer the necessary 3D structure and accurately estimate 6-DoF poses for all input cameras. To support this work, a new dataset called CrossGeo, comprising 278,000 images across 85 scenes, was also created for training and evaluation. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Enables more accurate 3D reconstruction for applications using diverse aerial and ground imagery.
RANK_REASON Academic paper detailing a new model and dataset for 3D reconstruction. [lever_c_demoted from research: ic=1 ai=1.0]