Cornell University researchers have developed a new computer-vision method, that enables machines to match real-world images with simplified building layouts like floor plans with much greater accuracy. To train and evaluate their approach, the team compiled a large dataset called C3, containing about 90,000 paired photos and floor plans across nearly 600 scenes, with detailed annotations of pixel matches and camera poses.
By reconstructing scenes in 3D from large internet photo collections and aligning them to publicly available architectural drawings, the dataset teaches models how real images relate to abstract representations. In tests, C3Po reduced matching errors by about 34% compared with earlier methods, suggesting that this multi-modal training could help future vision systems generalize across varied inputs and advance 3D computer vision research.
More information:
https://news.cornell.edu/stories/2025/12/computer-vision-connects-real-world-images-building-layouts