Automatic texturing for as-built building information model reconstruction from point cloud

• Main Author: Jin, Yuyang
• Other Authors: Cai Yiyu
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Building information modeling; Texture mapping; Point cloud
• Online Access: https://hdl.handle.net/10356/181702
• Institution: Nanyang Technological University

Most as-built Building Information Models (BIM) reconstructed today lack accurate texture information, appearing as white mold or using artificial texture. While manual texture mapping is feasible, it is highly labor-intensive, and no existing method can fully automate the restoration of texture information for as-built BIM based on the actual environment. Some research explored image-to-BIM registration techniques, but they relied on establishing feature correspondences between images, resulting in limited use cases. This paper proposes a fully automated approach to mapping accurate textures to as-built BIM using point cloud and video. Key challenges include converting precise positional point information to texture pixels and designing accurate texture maps for as-built BIM in Industry Foundation Classes (IFC) format. The methodology involves designing texture structure based on IFC Schema, projecting texture UV coordinates onto 3D points for sampling and filtering, training point cloud to 3D Gaussian splats for rendering, and customizing rendering projection methods. This automatic method has the following features: (1) Developed based on IFC format; (2) Using point cloud and video for texture creation, improving quality with Gaussian Splatting; (3) Thresholding technique is used for occlusion detection and removal, followed by machine learning methods to inpaint the removed parts; (4) Fully automated; (5) Applicable to as-built BIM reconstructed from point cloud either manually and automatically. In addition, quantified evaluations are performed with case studies and public domain datasets to ensure accurate mapping position with a mean error of within 4 pixels and a mean rotation error of within 0.06 degrees. Texture image qualities are also assessed, with both PSNR score and L1 score showing promising results, and demonstrating the proposed approach has the potential to facilitate as-built BIM texturing and realistic rendering.