Improving Edge Detection using Indoor Synthetic Data
Keywords
No Thumbnail Available
Authors
Issue Date
2020-01-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Synthetic data is often cheaper and easier to obtain than human-annotated data and can in some cases also
be much more accurate. Edge detection is one case where the use of fully convolutional neural networks has
led to an increase in performance in precision and recall but also introduced uncertainty of edge location due to
imprecisions in the edge labels. Standard measures for edge detection performance only re
ect the "correctness" of
edge predictions but not the precise localization of these edges. We propose a way to improve edge localization by
training a Bi-Directional Cascade Network on realistically rendered images of indoor scenes from the game engine
Unity. While doing this we introduce a precise formulation of edge detection, which is lacking in the literature.
We also investigate ways to bridge the gap between real image datasets and synthetic datasets using domain
randomization and CycleGANs for domain adaptation. The code and rendered data are online1.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen