Transfer Learning For Deep Convolutional Neural Network From Rgb To Ir Domain

Ligocki, Adam; Jelínek, Aleš

Transfer Learning For Deep Convolutional Neural Network From Rgb To Ir Domain

Files

366-eeict_1.pdf(2.31 MB)

Date

2020

Authors

Ligocki, Adam

Jelínek, Aleš

Publisher

Vysoké učení technické v Brně, Fakulta elektrotechniky a komunikačních technologií

Abstract

In this paper, we are presenting a proof of concept of our system for training of the YOLOv3 neural network for object detection of vehicles in thermal camera images. Our approach is unique in the way we are using a dataset containing a large number of synchronized range measurements as well as RGB and thermal images. We are using the existing YOLO toolkit to detect objects on the RGB images, we estimate detection distance by the LiDAR and later we reproject these detections into the IR image. In this way, we have created a large dataset of annotated thermal images that helped us to significantly improve the performance of the neural network at the IR domain.

Citation

Proceedings I of the 26st Conference STUDENT EEICT 2020: General papers. s. 366-370. ISBN 978-80-214-5867-3
https://conf.feec.vutbr.cz/eeict/EEICT2020

Document type

Peer-reviewed

Document version

Published version

Language of document

en

Document licence