A research team from the University of Cambridge’s Laing O'Rourke Centre has introduced InfraDiffusion, a zero-shot framework designed to restore infrastructure point clouds from sparse to dense by combining virtual camera projections with the strong generative capability of diffusion models.
Led by Dr Yixiong Jing, a postdoctoral researcher at the LOR Centre, together with fellow researchers Cheng Zhang, Haibing Wu, Dr Guangming Wang, Dr Olaf Wysocki, and Dr Brian Sheil, the team proposes a zero-shot pipeline, InfraDiffusion, capable of restoring noisy and incomplete depth maps of masonry infrastructure. Unlike previous approaches that primarily focused on component-level segmentation (such as arches, piers, etc.), InfraDiffusion enables zero-shot brick-level segmentation using the Segment Anything Model (SAM), paving the way toward scalable and automated inspection of masonry assets.
Tests conducted on four masonry infrastructure point clouds demonstrate significant improvements in segmentation accuracy by leveraging zero-shot SAM. These results represent a meaningful step forward in how engineers and researchers monitor and assess critical infrastructure.