Talks and Poster Presentations (with Proceedings-Entry):
D. Beran, N. Li, N. Pfeifer:
"Label errors in point cloud in training data for classification using machine learning";
Talk: Symposium GIS Ostrava 2020 UAV in Smart City and Smart Region,
Ostrava, Czechia and online;
- 2020-03-20; in: "Symposium GIS Ostrava 2020 UAV in Smart City and Smart Region",
One of the applications of ULS (UAV-borne laser scanning) lies in power line inspection. However, with LiDAR data (i.e. point clouds) comes the need for reliable automatic classification, also called semantic segmentation, of data which allows further analysis of gathered data. Vast number of possible methods for automatic classification of point clouds have been proposed and implemented, many of which depend on machine learning. Motivation for this research is the need for pre-classified data for training of machine learning models, specifically the impact of label accuracy/error in the pre-classified data used for machine learning classification. To find out what is the impact of error levels of labels on machine learning classification of power line point clouds we have used the method of Classification and regression trees (CART) using Opals software. During this research several tests were conducted with various levels and types of error in class labelling of training data and the results were compared with correctly labelled data to calculate confusion matrices and thus evaluate the impact of different error levels.
Created from the Publication Database of the Vienna University of Technology.