Authors:
Yubo Wang
1
;
Zhao Wang
1
;
Yuusuke Nakano
2
;
Katsuya Hasegawa
3
;
Hiroyuki Ishii
1
and
Jun Ohya
1
Affiliations:
1
Department of Modern Mechanical and Engineering, Waseda University, Tokyo, Japan
;
2
Network Service Systems Laboratories, NTT Corporation, Tokyo, Japan
;
3
Institute of Space and Astronautical Science, JAXA, Kanagawa, Japan
Keyword(s):
Geospatial Information Processing, Image Processing, Aerial Image Segmentation, Transformer Model, Feature Pyramid Network (FPN).
Abstract:
Unlike general semantic segmentation, aerial image segmentation has its own particular challenges, three
of the most prominent of which are great object scale variation, the scattering of multiple tiny objects in a
complex background and imbalance between foreground and background. Previous affinity learning-based
methods introduced intractable background noise but lost key-point information due to the additional interaction between different level features in their Feature Pyramid Network (FPN) like structure, which caused
inferior results.We argue that multi-scale information can be further exploited in each FPN level individually without cross-level interaction, then propose a Multi-scale Attention Cascade (MAC) model to leverage
spatial local contextual information by using multiple sized non-overlapping window self-attention module,
which mitigates the effect of complex and imbalanced background. Moreover, the multi-scale contextual cues
are propagated in a cascade manner to tac
kle the large scale variation problem while extracting further details.
Finally, a local channels attention is presented to achieve cross-channel interaction. Extensive experiments
verify the effectiveness of MAC and demonstrate that the performance of MAC surpasses those of the stateof-the-art approaches by +2.2 mIoU and +3.1 mFscore on iSAID dataset, by +2.97 mIoU on ISPRS Vaihingen
dataset. Code has been made available at https://github.com/EricBooob/Multi-scale-Attention-Cascade-forAerial-Image-Segmentation.
(More)