Unsupervised discovery of 3D structural elements for scanned indoor scenes Articles uri icon

publication date

  • February 2025

start page

  • 1

end page

  • 19

volume

  • 170

International Standard Serial Number (ISSN)

  • 1568-4946

Electronic International Standard Serial Number (EISSN)

  • 1872-9681

abstract

  • This paper addresses the growing demand for effective 3D sensing applications by presenting a comprehensive point cloud segmentation method developed for large indoor spaces. Our approach recognises the challenges associated with (un)ordered data and presents a robust algorithm capable of dealing with irregularities caused by measurement inaccuracies, e.g. occlusion, noise, outliers and discontinuous data transitions. The method uses a multi-step filtering approach that sequentially navigates through Gaussian map, distance space and regular grid representations. Connected component analysis, structural rules and assumptions guide the unsupervised clustering of structural elements (SEs), e.g. walls, ceilings and floors. The method is adaptable to various datasets, including joint 2D-3D datasets such as true RGB-D data. A colour metric is introduced to account for illumination effects during scanning and to ensure the generalisability of the method. The importance of detecting SEs lies in their role as input to deep neural networks, which improve the accuracy of SLAM algorithms and influence the quality of subsequent indoor residual object detection. This paper introduces density-based clustering of objects using colour similarity measures and low-level features to further refine the segmentation by eliminating outliers and improving the detection of sharp shapes. The proposed method represents a sophisticated and versatile solution that overcomes scene complexity and makes an important contribution to applications in scene understanding, SLAM and indoor object recognition.

subjects

  • Computer Science

keywords

  • point cloud; unsupervised search; statistical clustering; harmonic interpolation; structural elements; objects