Voxel-based 3D Point Cloud Semantic Segmentation: Unsupervised Geometric and Relationship Featuring vs Deep Learning Methods

Abstract

Automation in point cloud data processing is central in knowledge discovery within decision-making systems. The definition of relevant features is often key for segmentation and classification, with automated workflows presenting the main challenges. In this paper, we propose a voxel-based feature engineering that better characterize point clusters and provide strong support to supervised or unsupervised classification. We provide different feature generalization levels to permit interoperable frameworks. First, we recommend a shape-based feature set (SF1) that only leverages the raw X, Y, Z attributes of any point cloud. Afterwards, we derive relationship and topology between voxel entities to obtain a three-dimensional (3D) structural connectivity feature set (SF2). Finally, we provide a knowledge-based decision tree to permit infrastructure-related classification. We study SF1/SF2 synergy on a new semantic segmentation framework for the constitution of a higher semantic representation of point clouds in relevant clusters. Finally, we benchmark the approach against novel and best-performing deep-learning methods while using the full S3DIS dataset. We highlight good performances, easy-integration, and high F1-score (> 85%) for planar-dominant classes that are comparable to state-of-the-art deep learning.

Using Training Samples Retrieved from a Topographic Map and Unsupervised Segmentation for the Classification of Airborne Laser Scanning Data

“…The method, first, obtained well-segmented point clouds, then, the LDA was used to cluster the segment objects. For a more complex scene, Poux et al [2] derived two feature sets from the coordinate and the voxel entities. A knowledge-based decision tree was proposed using these features for unsupervised labeling of the point clouds.…”

“…Each level contains two scales, i.e., 16 neighbors and 32 neighbors. The search radius of 16 neighbors at each level are (2,4,8,16) meters. The search radius of 32 neighbors at each level are (4,8,16,32) meters.…”

“…Unsupervised methods seem to be an excellent solution to generate training samples. Some unsupervised methods [1,2] have been proposed to label indoor point clouds and have achieved good performances. However, for outdoor space, especially large scale urban scenes, it is difficult for unsupervised methods to classify objects correctly.…”

A Simple Semantic-Based Data Storage Layout for Querying Point Clouds

“…Moreover, SDBL via Python is specifically designed for querying via semantic classes, and there is no need for an additional indexing mechanism since indexing is handled through the operating system: the different folders contain different semantic classes. These features give SDBL via Python an advantage over file-based solutions for queries where semantics play a fundamental role, such as in the fields of autonomous driving and industrial robotic applications [78] or when using indoor navigation or BIM (Building Information Modeling) [79].…”

“…When the point is part of an object O 1 that rests on the ground (such as a table), it is classified at level 1 and, if the object O 1 acts as a host to another object O 2 , then that object O 2 is known as the first guest and is classified at a level immediately above, that is, level 2 in this case [25]. Such a scheme should prove very useful for indoor point cloud segmentation [79] and could be implemented using SDBL via Python for instance.…”