TY - GEN
T1 - Hybrid Perspective Mapping
T2 - 2021 IEEE International Intelligent Transportation Systems Conference, ITSC 2021
AU - Wang, Junbo
AU - Yang, Yi
AU - Pan, Miaoxin
AU - Zhang, Man
AU - Zhu, Minzhao
AU - Fu, Mengyin
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021/9/19
Y1 - 2021/9/19
N2 - Cross view image-based geo-localization aims to estimate global position of an image by matching the query image with the images in a geo-referenced image database. Cross-view image-based geo-localization is a potential supplement of GPS, but it's hard to matching cross-view image pairs because of the tremendous appearance differences. However, most of deep learning approaches match cross-view image pairs directly and ignore the inner relation between them. In this paper, we introduce our novel hybrid perspective mapping method to align ground-level image to aerial image by considering projection relation between them. Unlike other learning based method which generate corresponding aerial image by traning, our approach is totally geometry based and can be plugged to other network conveniently. And we propose our network based on hybrid perspective mapping. our network shows higher retrieval accuracy and powerful generalization ability on several public dataset. In addition, we also conduct cross-view image matching experiments on our own dataset and analyse the influence of spatial resolution and seasonal variation of aerial image on image matching.
AB - Cross view image-based geo-localization aims to estimate global position of an image by matching the query image with the images in a geo-referenced image database. Cross-view image-based geo-localization is a potential supplement of GPS, but it's hard to matching cross-view image pairs because of the tremendous appearance differences. However, most of deep learning approaches match cross-view image pairs directly and ignore the inner relation between them. In this paper, we introduce our novel hybrid perspective mapping method to align ground-level image to aerial image by considering projection relation between them. Unlike other learning based method which generate corresponding aerial image by traning, our approach is totally geometry based and can be plugged to other network conveniently. And we propose our network based on hybrid perspective mapping. our network shows higher retrieval accuracy and powerful generalization ability on several public dataset. In addition, we also conduct cross-view image matching experiments on our own dataset and analyse the influence of spatial resolution and seasonal variation of aerial image on image matching.
UR - http://www.scopus.com/inward/record.url?scp=85118440377&partnerID=8YFLogxK
U2 - 10.1109/ITSC48978.2021.9564573
DO - 10.1109/ITSC48978.2021.9564573
M3 - Conference contribution
AN - SCOPUS:85118440377
T3 - IEEE Conference on Intelligent Transportation Systems, Proceedings, ITSC
SP - 3040
EP - 3046
BT - 2021 IEEE International Intelligent Transportation Systems Conference, ITSC 2021
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 19 September 2021 through 22 September 2021
ER -