Name
Title | ||
|---|---|---|
How to Improve Deep Learning based Pavement Distress Detection while Minimizing Human Effort |
Abstract | ||
|---|---|---|
Aging public roads need frequent inspections in order to guarantee their permanent availability. Using deep neural networks, the process of detecting pavement distress can be automated to a high degree. However, evaluations show that they perform relatively poor on road images, that are significantly different from training data. Therefore, we show, how the performance can be improved with a human in the loop. The basic idea is to enlarge the training dataset. Luckily, many unlabeled road images from previous inspections are available. Nevertheless, annotating all of them is labor-intensive, and thus, not feasible. Since only diverse data enable an increase in performance, selecting the right subregions of the images for annotation is the key. To achieve this goal, we model the network's uncertainty and incorporate it for selecting new subregions. Our experiments show that we are able to improve the network's performance with only a fraction of data that would usually be necessary to get the same performance. |
Year | DOI | Venue |
|---|---|---|
2018 | 10.1109/COASE.2018.8560372 | 2018 IEEE 14th International Conference on Automation Science and Engineering (CASE) |
Keywords | Field | DocType |
public roads,deep neural networks,unlabeled road images,deep learning based pavement distress detection,human effort minimization | Training set,Distress,Annotation,Computer science,Artificial intelligence,Deep learning,Human-in-the-loop,Machine learning,Deep neural networks | Conference |
ISSN | ISBN | Citations |
2161-8070 | 978-1-5386-3594-0 | 1 |
PageRank | References | Authors |
0.41 | 0 | 4 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Daniel Seichter | 1 | 1 | 0.75 |
| Markus Eisenbach | 2 | 37 | 6.76 |
| Ronny Stricker | 3 | 54 | 5.18 |
| Horst-Michael Gross | 4 | 761 | 92.05 |