Integrating Active Learning and Contextually Guide for Semantic Labeling of LiDAR Point Cloud

To alleviate the difficulties in obtaining training data sets of 3D point clouds, an active learning (AL) framework is proposed to iteratively select a small portion of unlabeled points to query their labels, and creates a minimum manually-annotated training set. To handle the biased sampling problem caused by category imbalance and local similarities, a neighbor-consistency prior is used to conduct an unbiased sampling for selecting the value samples into the training set. Additionally, to reduce the number of categories used in labeling, a higher-order MRF containing a regional label cost term, is exploited to refine the labeling results.

1. Introduction

Currently, the traditional semantic labelling algorithm based on 3D point clouds requires a large number of labeled 3D point clouds for model training. Training data production needs an accurate category label for each 3D point in the 3D point cloud scene manually, but the method is at a considerable expense of time and money. According to statistics, marking one kilometer of point clouds data manually for dozens of hours. In addition, in some cases it is virtually impossible to accurately annotate the target in the scene. For example, there is a situation in which the scene data and its surroundings overlap each other, and the obtained point cloud data show that the objects are difficult to distinguish in the overlapping area. Therefore, traditional 3D point clouds semantic labelling algorithm is often restricted in the application process due to the factors that are difficult to obtain in training data.