Point-Denoise: Unsupervised outlier detection for 3D point clouds enhancement

Abstract

3D point cloud denoising is an increasingly demanding field as such type of data structure is getting more attention in perceiving the 3D environment for diverse applications. Despite their novelty, recently proposed solutions are still modest in terms of effectiveness and robustness, especially for scenes corrupted with a massive amount of noise. The encountered challenges are mainly due to the data acquisition process and the little-to-no knowledge of the statistical data distribution. In this paper, two promising unsupervised machine learning techniques are investigated, which are the Isolation Forest (If) and the Elliptic Envelope (EE). Each of these techniques detects noise using different philosophies. If uses a forest of iTrees; while EE uses a learned imaginary elliptic. The proposed solution, named Point-Denoise, tunes both techniques and fuses them at the decision-level. Although the solution simplicity, Point-Denoise reports superior results to state-of-the-art techniques. For evaluation purposes, both synthetic and real data are used. The chosen synthetic data is the ModelNet40 benchmark, which is augmented with a Gaussian and emulated 3D scanner noise with three different standard deviations: 0.5%, 1.0%, and 1.5% assessing the robustness of the proposed methodology. Meanwhile, the real data is collected from the Qatar University campus. Considering that a massive amount of noise already corrupts real data at acquisition time, no additional noise is augmented. Point-Denoise outperforms state-of-the-art solutions (i.e., traditional filtering, supervised, and unsupervised learning techniques) by attaining a 0.24 distance error and achieving a 48.93% enhancement.