Feature Relevance Analysis for 3d Point Cloud Classification Using Deep Learning

Abstract

<p><strong>Abstract.</strong> 3D point clouds acquired by laser scanning and other techniques are difficult to interpret because of their irregular structure. To make sense of this data and to allow for the derivation of useful information, a segmentation of the points in groups, units, or classes fit for the specific use case is required. In this paper, we present a non-end-to-end deep learning classifier for 3D point clouds using multiple sets of input features and compare it with an implementation of the state-of-the-art deep learning framework PointNet++. We first start by extracting features derived from the local normal vector (normal vectors, eigenvalues, and eigenvectors) from the point cloud, and study the result of classification for different local search radii. We extract additional features related to spatial point distribution and use them together with the normal vector-based features. We find that the classification accuracy improves by up to 33% as we include normal vector features with multiple search radii and features related to spatial point distribution. Our method achieves a mean Intersection over Union (mIoU) of 94% outperforming PointNet++’s Multi Scale Grouping by up to 12%. The study presents the importance of multiple search radii for different point cloud features for classification in an urban 3D point cloud scene acquired by terrestrial laser scanning.</p>

An Efficient Deep Learning Approach for Ground Point Filtering in Aerial Laser Scanning Point Clouds

“…However, feature-based approaches do depend on a clear understanding of the features and their relation to the underlying problem. Feature-based DL has been used in point cloud classification (Zhang et al, 2018;Kumar et al, 2019). Kumar et al (2019) developed a multi-scale non end-to-end method for classification of terrestrial laser scanning (TLS) data.…”

“…Feature-based DL has been used in point cloud classification (Zhang et al, 2018;Kumar et al, 2019). Kumar et al (2019) developed a multi-scale non end-to-end method for classification of terrestrial laser scanning (TLS) data. In (Kumar et al 2019), because of multi-scaling, four normal vector based features (normal, plane-fit quality, eigenvalues and eigenvectors) are used a number of times.…”

“…Kumar et al (2019) developed a multi-scale non end-to-end method for classification of terrestrial laser scanning (TLS) data. In (Kumar et al 2019), because of multi-scaling, four normal vector based features (normal, plane-fit quality, eigenvalues and eigenvectors) are used a number of times. The authors claim that their method performs better than PointNet++ (Qi et al, 2017b), which is a multi-scale approach.…”

“…The authors claim that their method performs better than PointNet++ (Qi et al, 2017b), which is a multi-scale approach. Since, the method (Kumar et al, 2019) requires local features from multiple (radii) sizes of neighborhoods, it increases the computational burden. Additionally, we do not see any specific guideline on how the authors (Kumar et al, 2019) fix the size of the different radii to get the local neighborhood and corresponding features of multiscale interest points.…”

Efficient Loading and Visualization of Massive Feature-Rich Point Clouds Without Hierarchical Acceleration Structures

“…Nevertheless, attributes of different categories were also used in the overall process again. (Kumar et al, 2019), who apply neural networks for semantic segmentation, also rely on features computed in neighborhoods of different size. No hand-crafted features are required in deep learning (Liu et al, 2019), but it may be of interest to study the learned features, e.g.…”