SEEING THE BIGGER PICTURE: ENABLING LARGE CONTEXT WINDOWS IN NEURAL NETWORKS BY COMBINING MULTIPLE ZOOM LEVELS

Konrad Heidler; Lichao Mou; Xiao Xiang Zhu

doi:10.1109/IGARSS47720.2021.9554434

SEEING THE BIGGER PICTURE: ENABLING LARGE CONTEXT WINDOWS IN NEURAL NETWORKS BY COMBINING MULTIPLE ZOOM LEVELS

Konrad Heidler
, Lichao Mou
, Xiao Xiang Zhu

Chair of Data Science in Earth Observation

Technical University of Munich
Deutsches Zentrum für Luft- und Raumfahrt (DLR)

Research output: Contribution to conference › Paper › peer-review

3 Scopus citations

Abstract

When adopting deep learning methods for remote sensing applications, the data usually needs to be cut into patches due to hardware limitations. Clearly, this practice discards a lot of contextual information as the model’s information is limited to imagery from the given patch. We propose a memory-efficient way around this limitation by using multiple patches of varying spatial extents on different resolution levels. Finally, this new approach is evaluated for the task of automated sea ice charting, where the added contextual information is shown to be beneficial to model performance.

Original language	English
Pages	3033-3036
Number of pages	4
DOIs	https://doi.org/10.1109/IGARSS47720.2021.9554434
State	Published - 2021
Event	2021 IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021 - Brussels, Belgium Duration: 12 Jul 2021 → 16 Jul 2021

Conference

Conference	2021 IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021
Country/Territory	Belgium
City	Brussels
Period	12/07/21 → 16/07/21

Keywords

Image segmentation
Multiresolution
Sea ice
Synthetic aperture radar

Access to Document

10.1109/IGARSS47720.2021.9554434

Cite this

@conference{2c7e09d239864c52bf755203bf312091,

title = "SEEING THE BIGGER PICTURE: ENABLING LARGE CONTEXT WINDOWS IN NEURAL NETWORKS BY COMBINING MULTIPLE ZOOM LEVELS",

abstract = "When adopting deep learning methods for remote sensing applications, the data usually needs to be cut into patches due to hardware limitations. Clearly, this practice discards a lot of contextual information as the model{\textquoteright}s information is limited to imagery from the given patch. We propose a memory-efficient way around this limitation by using multiple patches of varying spatial extents on different resolution levels. Finally, this new approach is evaluated for the task of automated sea ice charting, where the added contextual information is shown to be beneficial to model performance.",

keywords = "Image segmentation, Multiresolution, Sea ice, Synthetic aperture radar",

author = "Konrad Heidler and Lichao Mou and Zhu, \{Xiao Xiang\}",

note = "Publisher Copyright: {\textcopyright} 2021 IEEE; 2021 IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021 ; Conference date: 12-07-2021 Through 16-07-2021",

year = "2021",

doi = "10.1109/IGARSS47720.2021.9554434",

language = "English",

pages = "3033--3036",

}

TY - CONF

T1 - SEEING THE BIGGER PICTURE

T2 - 2021 IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2021

AU - Heidler, Konrad

AU - Mou, Lichao

AU - Zhu, Xiao Xiang

PY - 2021

Y1 - 2021

N2 - When adopting deep learning methods for remote sensing applications, the data usually needs to be cut into patches due to hardware limitations. Clearly, this practice discards a lot of contextual information as the model’s information is limited to imagery from the given patch. We propose a memory-efficient way around this limitation by using multiple patches of varying spatial extents on different resolution levels. Finally, this new approach is evaluated for the task of automated sea ice charting, where the added contextual information is shown to be beneficial to model performance.

AB - When adopting deep learning methods for remote sensing applications, the data usually needs to be cut into patches due to hardware limitations. Clearly, this practice discards a lot of contextual information as the model’s information is limited to imagery from the given patch. We propose a memory-efficient way around this limitation by using multiple patches of varying spatial extents on different resolution levels. Finally, this new approach is evaluated for the task of automated sea ice charting, where the added contextual information is shown to be beneficial to model performance.

KW - Image segmentation

KW - Multiresolution

KW - Sea ice

KW - Synthetic aperture radar

UR - https://www.scopus.com/pages/publications/85128002076

U2 - 10.1109/IGARSS47720.2021.9554434

DO - 10.1109/IGARSS47720.2021.9554434

M3 - Paper

AN - SCOPUS:85128002076

SP - 3033

EP - 3036

Y2 - 12 July 2021 through 16 July 2021

ER -

SEEING THE BIGGER PICTURE: ENABLING LARGE CONTEXT WINDOWS IN NEURAL NETWORKS BY COMBINING MULTIPLE ZOOM LEVELS

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this