Deep learning based dense matching for aerial remote sensing images
This work studied that the application of deep learning based stereo methods in aerial remote sensing images, including its performance evaluation, the comparison with classical methods and generalization ability estimation.Three convolution neural networks are applied, MC-CNN(matching cost convolut...
Main Authors: | , |
---|---|
Format: | Article |
Language: | zho |
Published: |
Surveying and Mapping Press
2019-09-01
|
Series: | Acta Geodaetica et Cartographica Sinica |
Subjects: | |
Online Access: | http://html.rhhz.net/CHXB/html/2019-9-1141.htm |
id |
doaj-527619fd618f46f9a0048a3b4cc86a84 |
---|---|
record_format |
Article |
spelling |
doaj-527619fd618f46f9a0048a3b4cc86a842020-11-25T02:25:57ZzhoSurveying and Mapping PressActa Geodaetica et Cartographica Sinica1001-15951001-15952019-09-014891141115010.11947/j.AGCS.2019.201802472019090247Deep learning based dense matching for aerial remote sensing imagesLIU Jin0JI Shunping1School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, ChinaSchool of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, ChinaThis work studied that the application of deep learning based stereo methods in aerial remote sensing images, including its performance evaluation, the comparison with classical methods and generalization ability estimation.Three convolution neural networks are applied, MC-CNN(matching cost convolutional neural network), GC-Net(geometry and context network) and DispNet(disparity estimation network), on aerial stereo image pairs. The results are compared with SGM (semi-global matching) and a commercial software SURE. Secondly, the generalization ability of the MC-CNN and GC-Net are evaluated with models pretrained on other datasets. Finally, fine tuning on a small number of target training data with pretrained models are compared to direct training. Three sets of aerial images and two open-source street data sets are used for test. Experiments show that:firstly, deep learning methods perform slightly better than traditional methods; secondly, both GC-Net and MC-CNN have demonstrated good generalization ability, and can get satisfactory 3PE (3-pixel-error) results on aerial images using a model pretrained on available stereo benchmarks; thirdly, when the training samples in target dataset are insufficient, the strategy of fine-tuning on a pretrained model can improve the effect of direct training.http://html.rhhz.net/CHXB/html/2019-9-1141.htmstereo matchingdense matchingaerial imagesconvolutional neural networkdeep learning |
collection |
DOAJ |
language |
zho |
format |
Article |
sources |
DOAJ |
author |
LIU Jin JI Shunping |
spellingShingle |
LIU Jin JI Shunping Deep learning based dense matching for aerial remote sensing images Acta Geodaetica et Cartographica Sinica stereo matching dense matching aerial images convolutional neural network deep learning |
author_facet |
LIU Jin JI Shunping |
author_sort |
LIU Jin |
title |
Deep learning based dense matching for aerial remote sensing images |
title_short |
Deep learning based dense matching for aerial remote sensing images |
title_full |
Deep learning based dense matching for aerial remote sensing images |
title_fullStr |
Deep learning based dense matching for aerial remote sensing images |
title_full_unstemmed |
Deep learning based dense matching for aerial remote sensing images |
title_sort |
deep learning based dense matching for aerial remote sensing images |
publisher |
Surveying and Mapping Press |
series |
Acta Geodaetica et Cartographica Sinica |
issn |
1001-1595 1001-1595 |
publishDate |
2019-09-01 |
description |
This work studied that the application of deep learning based stereo methods in aerial remote sensing images, including its performance evaluation, the comparison with classical methods and generalization ability estimation.Three convolution neural networks are applied, MC-CNN(matching cost convolutional neural network), GC-Net(geometry and context network) and DispNet(disparity estimation network), on aerial stereo image pairs. The results are compared with SGM (semi-global matching) and a commercial software SURE. Secondly, the generalization ability of the MC-CNN and GC-Net are evaluated with models pretrained on other datasets. Finally, fine tuning on a small number of target training data with pretrained models are compared to direct training. Three sets of aerial images and two open-source street data sets are used for test. Experiments show that:firstly, deep learning methods perform slightly better than traditional methods; secondly, both GC-Net and MC-CNN have demonstrated good generalization ability, and can get satisfactory 3PE (3-pixel-error) results on aerial images using a model pretrained on available stereo benchmarks; thirdly, when the training samples in target dataset are insufficient, the strategy of fine-tuning on a pretrained model can improve the effect of direct training. |
topic |
stereo matching dense matching aerial images convolutional neural network deep learning |
url |
http://html.rhhz.net/CHXB/html/2019-9-1141.htm |
work_keys_str_mv |
AT liujin deeplearningbaseddensematchingforaerialremotesensingimages AT jishunping deeplearningbaseddensematchingforaerialremotesensingimages |
_version_ |
1724849297320574976 |