Instance segmentation of front doors in mobile mapping system images

Klimkowska, Anna Maria (2020) Instance segmentation of front doors in mobile mapping system images. MRes thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (11MB)

Abstract

Scene understanding, through instance segmentation, is one of the most challenging and difficult tasks in computer vision. Instance segmentation is a combination of object detection and semantic segmentation, resulting in assigning each pixel in an image to a given class. This research aims to investigate the capabilities of the Mask R-CNN architecture to detect front doors from images delivered from a Mobile Mapping System. It was achieved by conducting a set of experiments focusing on training the Mask R-CNN with different combinations of training datasets. Front doors are important objects of interest in this work due to their significance in various applications like enriching information of building Level of Details, or recognising door locations for emergency services and people with disabilities. The data used in the training and testing phases included images collected from different locations (cities in the UK and Poland) and of various types (mobile phone, Mobile Mapping System, RGB and gray images). The results of the analysis suggest that increasing the size of the training dataset improves model capabilities for front door detection. Moreover, the analysis of the performance of models trained on datasets from different locations shows that models can be used to detect front door instances regardless of the origin of the training and test datasets. However, despite the satisfactory results, a comparative analysis of the Mask R-CNN and YOLOv3 approaches showed the superiority of the YOLOv3 over Mask R-CNN in the overall model performance. The best data combination used to train Mask R-CNN resulted in an F1-score equal to 0.7 which increased to 0.74 through the implementation of a post-processing step based on a heightto-width ratio threshold. The research carried out in this project shows the potential of the examined method for front door detection from Mobile Mapping Systems.

Item Type: Thesis (University of Nottingham only) (MRes)
Supervisors: Grebby, Stephen
Marsh, Stuart
Keywords: MMS, deep learning, object detection, camera
Subjects: T Technology > TA Engineering (General). Civil engineering (General) > TA1501 Applied optics. Phonics
Faculties/Schools: UK Campuses > Faculty of Engineering > Department of Civil Engineering
Item ID: 63806
Depositing User: Klimkowska, Anna
Date Deposited: 12 Jan 2021 14:32
Last Modified: 12 Jan 2021 14:45
URI: http://eprints.nottingham.ac.uk/id/eprint/63806

Actions (Archive Staff Only)

Edit View Edit View