Authors:
Houssem-Eddine Benseddik
1
;
Ariane Herbulot
1
;
2
and
Michel Devy
1
Affiliations:
1
LAAS, CNRS, Toulouse, France
;
2
Univ. de Toulouse, UPS, LAAS, F-31400 Toulouse, France
Keyword(s):
Airplane Detection, YOLOv4, Deep Learning, Domain Randomization, Object Detection, Synthetic Data.
Abstract:
This paper proposes a novel approach to generate a synthetic dataset through domain randomization, to address the problem of real-time airplane detection on airport zones with high accuracy. Most solutions have been employed and developed across satellite images with deep learning techniques. Our approach specifically targets airplane detection on complex airport environment using deep learning approach as YOLOv4. To improve training, a large amount of annotated training data are required for good performance. To address this issue, this study proposes the use of synthetic training data. There is however a large performance gap between methods trained on real and synthetic data. This paper introduces a new method, which bridges this gap based upon Domain Randomization. The approach is evaluated on bounding box detection of airplanes on the FGVC-Aircraft dataset.