University of Twente Student Theses

Login

Learning Behaviour of Sparse Point-Voxel Convolution: Semantic Segmentation of Railway LiDAR scans

Werf, Jasper van der (2023) Learning Behaviour of Sparse Point-Voxel Convolution: Semantic Segmentation of Railway LiDAR scans.

[img] PDF
633kB
Abstract:Supervised learning approaches require the creation of big datasets without upfront knowledge of the performance on these datasets. The high cost associated with such datasets highlights importance of being able to make performance estimates for full datasets through the analysis of learning behaviour on smaller datasets. This work analyzes learning behaviour with respect to variations in dataset size through a comparison of per-class intersection-over-union (IoU) against point- and scene-count in training data. In total, SPVConv models are trained for semantic segmentation of railways on various dataset-sizes. Linear regressions are extrapolated for the upward-trending performance on test data against the downward-trending performance on training data for both scene- and point-count, resulting in per-class predictions of IoU at their intersections. This work shows that some of the seen variations in IoU between the classes is very likely caused by a big class-imbalance in the dataset; this correlation is seen on limited data but also holds as the amount of data increases. In addition to the class imbalance, there are additional class-intrinsic factors that impact learning rate and IoU, shown through differences in slope for the various classes.
Item Type:Essay (Bachelor)
Clients:
Strukton Rail, Enschede, Netherlands
Saxion University of Applied Sciences, Enschede, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science BSc (56964)
Awards:Best Paper Award
Link to this item:https://purl.utwente.nl/essays/96106
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page