Enhancing data informativeness in deep-learning based, point cloud semantic segmentation of industrial scenes
Year: 2023
Type: dissertation
Abstract: Thanks to current technologies, reconstructing our environments as a 3D spaces is easier. However, transforming existing industrial facilities to Building Information Model (BIM) via Computer Assisted Drawing (CAD)software is still a tedious task. The goal of this thesis is to improve performance of point cloud semantic segmentation applied to an industrial setting. The literature review showed that current state of the art methods achieve high segmentation results on publicly available dataset but no such dataset exist for an industrial setting. A first study of available data allowed us to show the specificity and difficulties in working with industrial data. These peculiarities, combined with a lack in data quantity, heads this thesis towards finding ways to use available data more effectively. To solve this task, two ideas were explored. First, asynthetic data generation method was created and the relationship between synthetic data realism and segmentation performanceis studied. Secondly, different data transformation to be applied before the segmentation network input are studied and the ability of some of them to increase segmentation performance is shown.
Source: HAL
Author: Cazorla, Romain
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6989087826