A Method for Vehicle Detection in High-Resolution Satellite Images that Uses a Region-Based Object Detector and Unsupervised Domain Adaptation

Yohei Koga,Ryosuke Shibasaki,Hiroyuki Miyazaki

doi:10.3390/rs12030575

Abstract

Recently, object detectors based on deep learning have become widely used for vehicle detection and contributed to drastic improvement in performance measures. However, deep learning requires much training data, and detection performance notably degrades when the target area of vehicle detection (the target domain) is different from the training data (the source domain). To address this problem, we propose an unsupervised domain adaptation (DA) method that does not require labeled training data, and thus can maintain detection performance in the target domain at a low cost. We applied Correlation alignment (CORAL) DA and adversarial DA to our region-based vehicle detector and improved the detection accuracy by over 10% in the target domain. We further improved adversarial DA by utilizing the reconstruction loss to facilitate learning semantic features. Our proposed method achieved slightly better performance than the accuracy achieved with the labeled training data of the target domain. We demonstrated that our improved DA method could achieve almost the same level of accuracy at a lower cost than non-DA methods with a sufficient amount of labeled training data of the target domain.

Highlights

Vehicle detection methods have been drastically improving due to the use of deep learning
We explored improving the accuracy in a region of interest (RoI) using a vehicle detector trained on labeled data, which corresponds to a domain adaptation (DA) problem
Our contribution is twofold: (1) we adapted two DA methods that were originally proposed for image classification networks to an object detection network and applied Correlation alignment (CORAL) DA and adversarial DA to our region-based vehicle detector, and (2) we propose a novel DA method of combining adversarial DA and image reconstruction and demonstrate its effectiveness

Summary

Introduction

Vehicle detection methods have been drastically improving due to the use of deep learning. The deep learning method requires adequate training to achieve high levels of accuracy. Several large-scale open datasets for vehicle detection have been proposed, the region of interest (RoI) in practice (the “target domain”) is often different from that in the datasets (the “source domain”). In such cases, vehicle detection performance in the target domain is notably low because of differences in image features between the source and target domains. The method of Tang et al [1] could achieve high accuracy in the source domain (Munich dataset), that method could not address the image feature difference between the source and the target (collected vehicle dataset) domain; the accuracy in the target domain was notably lower than the source domain. To achieve a useful level of accuracy, additional training data are required in the target domain, but adding this step is quite costly

Methods

Findings

Discussion

Conclusion