Multiscale Point Cloud Optimization for SfM Reconstruction

The 3D scanning and reconstruction technologies for augmented reality, autonomous vehicles, remote sensing, GIS, object recognition and localization, are often impaired by a critical fundamental problem: noise. Noise is even more problematic in 3D models generated by SfM, or Structure from Motion. Here, the outlier points have an adverse effect on the processing and application of the generated point clouds that are often the basis for 3D feature detection, localization and navigation algorithms. In general, visualization of 3D environments, navigation, and volumetric medical image segmentation present numerous challenges when noisy outliers interfere with the surface delimiters of the scanned objects. In this paper, we propose an effective strategy to filter noisy points generated in the process of SfM reconstruction. We formulate a novel approach based on an adaptive moving least squares (MLS) to optimize the geometric structure of a typical 3D indoor scene model. Different from other existing adaptive MLS, our method considers the adverse interactions between the neighboring non-continuous model components. The effectiveness and the performance of our approach is demonstrated on extended indoor scene models generated from 3D point clouds based on SfM.