pith. sign in

arxiv: 2605.26381 · v1 · pith:II2427KEnew · submitted 2026-05-25 · 💻 cs.CV

Multi-Modal Building Inspection via Perceiver IO Fusion of Satellite and Street-Level Imagery

Niels Sombekke , Rob G.J. Wijnhoven , Martin R. Oswald This is my paper

Pith reviewed 2026-06-29 22:14 UTC · model grok-4.3

classification 💻 cs.CV
keywords multi-modal fusionPerceiver IOsatellite imagerystreet-level imagerybuilding classificationroof materialsDINOv2computer vision
0
0 comments X

The pith

Perceiver IO fusion of satellite and street-level images improves classification of roof attributes best seen from the ground.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a framework that feeds tokens from a shared DINOv2 backbone into a Perceiver IO model to combine satellite views with a variable number of street-level photos for multi-label prediction of roof materials and elements. A new dataset covers 32,135 buildings across ten countries, and the authors test four masking strategies, finding that appending a building footprint mask as a fourth RGB-M channel helps the model focus on the target structure. The Perceiver IO approach outperforms other fusion methods and delivers clear per-class gains on street-visible features such as slate roofs, while a satellite-only baseline keeps a small edge on classes that are mainly visible from above.

Core claim

A Perceiver IO architecture that ingests spatial patch tokens from a shared DINOv2 backbone, naturally accommodates an arbitrary number of street-level views per building, and jointly predicts multi-label roof element and material classes improves over alternative fusion strategies; it produces substantial per-class gains for attributes visible from street level while the satellite-only baseline retains a slight advantage in macro-averaged mAP for classes predominantly visible from above.

What carries the argument

Perceiver IO model that processes variable-length sets of DINOv2 patch tokens from satellite and street-level images and outputs multi-label classifications without padding or fixed-size pooling.

If this is right

  • RGB-M masking that adds the building footprint as a fourth channel outperforms hard cropping on both modalities.
  • Per-class average precision rises by 11.3 points for slate and 1.3 points for dormers when street-level views are fused.
  • The architecture scales to heterogeneous inputs and multiple output tasks without requiring fixed view counts.
  • Satellite-only remains competitive or superior for roof attributes that are mainly visible from above.
  • The same fusion strategy can be applied to other building-inspection tasks that combine overhead and ground-level imagery.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The variable-view handling could be tested on tasks such as facade damage detection where the number of available street photos also varies.
  • The RGB-M masking prior might transfer to other multi-modal settings that need soft spatial focus without explicit cropping.
  • The ten-country dataset could serve as a starting benchmark for studying domain shift when models are trained on one region and tested on another.
  • Extending the fusion to include temporal sequences of street views could reveal whether change detection benefits from the same Perceiver IO design.

Load-bearing premise

The constructed dataset of 32,135 buildings with paired satellite and street-level images and labels supplies an unbiased test of multi-modal performance without significant labeling errors or domain shift.

What would settle it

Retraining and evaluating the same Perceiver IO model on a fresh set of buildings from additional countries or with independently verified labels, then checking whether the reported per-class gains for street-visible attributes disappear or reverse.

Figures

Figures reproduced from arXiv: 2605.26381 by Martin R. Oswald, Niels Sombekke, Rob G.J. Wijnhoven.

Figure 1
Figure 1. Figure 1: Example input data for a building: satellite image (left) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset statistics: building and segment counts by coun [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of street-level images per building (left) and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Satellite imagery challenges: no issues, low reso [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Street-level imagery challenges: no issues, occlusion by [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Examples of street-level occlusion patterns. Vegetation, [PITH_FULL_IMAGE:figures/full_fig_p003_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Mask refinement limitation: sparse vegetation in the [PITH_FULL_IMAGE:figures/full_fig_p004_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Street-level mask generation pipeline. From left to right: [PITH_FULL_IMAGE:figures/full_fig_p004_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Uni-modal satellite model. The RGB-M input (4 chan [PITH_FULL_IMAGE:figures/full_fig_p005_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Uni-modal street-level model. Each street-level image [PITH_FULL_IMAGE:figures/full_fig_p005_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Concatenation fusion baseline. Satellite and aggregated [PITH_FULL_IMAGE:figures/full_fig_p005_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Feature Vector Transformer fusion. Individual feature [PITH_FULL_IMAGE:figures/full_fig_p005_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Perceiver IO fusion architecture. A satellite image and [PITH_FULL_IMAGE:figures/full_fig_p006_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Grad-CAM visualizations for satellite masking strate [PITH_FULL_IMAGE:figures/full_fig_p007_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: Perceiver IO sensitivity to latent configuration ( [PITH_FULL_IMAGE:figures/full_fig_p007_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Grad-CAM visualizations for satellite imagery across [PITH_FULL_IMAGE:figures/full_fig_p008_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Grad-CAM visualizations for street-level imagery with [PITH_FULL_IMAGE:figures/full_fig_p008_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Perceiver IO attention rollout visualizations for three buildings. Each row shows the satellite image (left, with attention overlay) [PITH_FULL_IMAGE:figures/full_fig_p009_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Perceiver IO attention rollout for chimney detection— [PITH_FULL_IMAGE:figures/full_fig_p009_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Challenging classes: Aluminium (top), Corrugated [PITH_FULL_IMAGE:figures/full_fig_p010_22.png] view at source ↗
read the original abstract

We present a multi-modal classification framework that fuses satellite and street-level imagery through a Perceiver IO architecture operating on spatial patch tokens from a shared DINOv2 backbone. The design naturally handles a variable number of street-level views per building without padding or fixed-size pooling, and jointly predicts multi-label roof element and roof material classes. We construct a large-scale dataset of 32,135 buildings (61,672 segments) spanning ten countries, pairing satellite images with up to eight street-level views per segment and evaluating four masking strategies for isolating the target building. We propose an RGB-M masking strategy that appends the building footprint mask as a fourth input channel, providing a soft spatial prior that outperforms hard cropping across both modalities. The Perceiver IO fusion model improves over all other fusion strategies and yields substantial per-class gains for attributes visible from street level (e.g., +11.3 AP for slate, +1.3 AP for dormers), though the satellite-only baseline retains a slight advantage in macro-averaged mAP for classes that are predominantly visible from above. These results establish a scalable, flexible architecture for multi-modal building inspection that can accommodate heterogeneous inputs and multiple output tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper presents a Perceiver IO architecture that fuses spatial patch tokens from a shared DINOv2 backbone applied to satellite and variable numbers of street-level images for multi-label classification of roof elements and roof materials. It introduces a dataset of 32,135 buildings (61,672 segments) across ten countries, evaluates four masking strategies including a proposed RGB-M channel, and reports that the Perceiver IO fusion model outperforms other fusion baselines with notable per-class AP gains for street-visible attributes (e.g., +11.3 AP for slate) while satellite-only remains competitive on macro mAP for overhead-visible classes.

Significance. If the empirical results are reliable, the work demonstrates a flexible multi-modal architecture that naturally accommodates heterogeneous inputs without fixed pooling or padding, together with a large-scale multi-country dataset that could support further research on building attribute prediction. The concrete per-class gains for street-visible attributes and the comparison across masking strategies provide a useful empirical baseline for multi-modal fusion in remote sensing.

major comments (2)
  1. [Dataset section] Dataset section (abstract and § on data construction): the 32,135-building dataset is central to all reported AP improvements, yet the manuscript supplies no description of the labeling process (expert, crowdsourced, or automated), inter-annotator agreement, or cross-country consistency checks. Without these details it is impossible to rule out systematic annotation bias or noise that could inflate the claimed +11.3 AP gain for slate and the fusion-vs-baseline comparisons.
  2. [Results and training sections] Experimental protocol (results and training sections): the abstract reports concrete AP numbers and masking-strategy comparisons, but no information is given on training hyperparameters, data splits, optimizer settings, or statistical testing of the observed differences. This absence directly limits verification of the central claim that Perceiver IO fusion is superior.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will revise the manuscript to incorporate the requested details where they are available.

read point-by-point responses

  1. Referee: [Dataset section] Dataset section (abstract and § on data construction): the 32,135-building dataset is central to all reported AP improvements, yet the manuscript supplies no description of the labeling process (expert, crowdsourced, or automated), inter-annotator agreement, or cross-country consistency checks. Without these details it is impossible to rule out systematic annotation bias or noise that could inflate the claimed +11.3 AP gain for slate and the fusion-vs-baseline comparisons.

    Authors: We agree that the absence of labeling methodology details limits assessment of dataset quality. In the revised manuscript we will add a dedicated paragraph in the data construction section describing the annotation process (including whether labels were expert-annotated or otherwise obtained), any inter-annotator agreement figures that were computed, and the procedures used to ensure cross-country label consistency. If certain agreement statistics were not collected during dataset creation, we will explicitly state this. revision: yes

  2. Referee: [Results and training sections] Experimental protocol (results and training sections): the abstract reports concrete AP numbers and masking-strategy comparisons, but no information is given on training hyperparameters, data splits, optimizer settings, or statistical testing of the observed differences. This absence directly limits verification of the central claim that Perceiver IO fusion is superior.

    Authors: We acknowledge that full experimental details are required for reproducibility and verification. The revised manuscript will include a new subsection under the training or experimental protocol heading that specifies the train/validation/test splits, optimizer, learning-rate schedule, batch size, number of epochs, and any other hyperparameters. We will also report whether statistical significance tests were performed on the AP differences and, if so, the results; if no such tests were conducted we will state this explicitly. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation on newly constructed dataset

full rationale

The paper reports experimental results from training and evaluating a Perceiver IO fusion model on a newly assembled dataset of 32,135 buildings. No equations, predictions, or uniqueness claims are present that reduce reported gains (e.g., +11.3 AP) to quantities defined by fitted parameters, self-citations, or ansatzes imported from prior author work. All comparisons are external benchmarks against other fusion strategies on held-out data, making the derivation chain self-contained and non-circular.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the quality and representativeness of the newly constructed dataset and on the assumption that DINOv2 patch tokens transfer effectively to both modalities without domain-specific adaptation.

free parameters (1)
  • maximum number of street-level views
    Model accepts variable inputs up to eight views per segment; the choice of eight is a design decision that affects training and evaluation.
axioms (1)
  • domain assumption DINOv2 provides suitable spatial patch tokens for both satellite and street-level imagery
    Shared backbone is used without reported fine-tuning or modality-specific justification in the abstract.

pith-pipeline@v0.9.1-grok · 5749 in / 1311 out tokens · 35970 ms · 2026-06-29T22:14:53.347464+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    Street view imagery in urban analytics and gis: A review.Landscape and Urban Planning, 215:104217, 2021

    Filip Biljecki and Koichi Ito. Street view imagery in urban analytics and gis: A review.Landscape and Urban Planning, 215:104217, 2021. 2

    2021

  2. [2]

    Integrating aerial and street view images for urban land use classification.Re- mote Sensing, 10(10):1553, 2018

    Rui Cao, Jiasong Zhu, Wei Tu, Qingquan Li, Jinzhou Cao, Bozhi Liu, Qian Zhang, and Guoping Qiu. Integrating aerial and street view images for urban land use classification.Re- mote Sensing, 10(10):1553, 2018. 1, 2

    2018

  3. [3]

    Boan Chen, Quanlong Feng, Bowen Niu, Fengqin Yan, Bingbo Gao, Jianyu Yang, Jianhua Gong, and Jiantao Liu. Multi-modal fusion of satellite and street-view images for ur- ban village classification based on a dual-branch deep neural network.International Journal of Applied Earth Observa- tion and Geoinformation, 109:102794, 2022. 2

    2022

  4. [4]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 3

    2016

  5. [5]

    Energy performance of build- ings directive.https : / / energy

    European Commission. Energy performance of build- ings directive.https : / / energy . ec . europa . eu / topics / energy - efficiency / energy - efficient- buildings/energy- performance- buildings-directive_en, 2024. 1

    2024

  6. [6]

    Runyu Fan, Jun Li, Fengpeng Li, Wei Han, and Lizhe Wang. Multilevel spatial-channel feature fusion network for urban village classification by fusing satellite and streetview im- ages.IEEE Transactions on Geoscience and Remote Sens- ing, 60:1–13, 2022. 2

    2022

  7. [7]

    Automatic detection of building ty- pology using deep learning methods on street level images

    Daniela Gonzalez, Diego Rueda-Plata, Ana B Acevedo, Juan C Duque, Ra ´ul Ramos-Poll ´an, Alejandro Betancourt, and Sebastian Garc ´ıa. Automatic detection of building ty- pology using deep learning methods on street level images. Building and Environment, 177:106805, 2020. 2

    2020

  8. [8]

    Fusion of satellite and street view data for urban traffic accident hotspot identi- fication.International Journal of Applied Earth Observation and Geoinformation, 130:103853, 2024

    Wentong Guo, Cheng Xu, and Sheng Jin. Fusion of satellite and street view data for urban traffic accident hotspot identi- fication.International Journal of Applied Earth Observation and Geoinformation, 130:103853, 2024. 2

    2024

  9. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 4

    2016

  10. [10]

    Urban neighbourhood environ- ment assessment based on street view image processing: A review of research trends.Environmental Challenges, 4: 100090, 2021

    Nan He and Guanghao Li. Urban neighbourhood environ- ment assessment based on street view image processing: A review of research trends.Environmental Challenges, 4: 100090, 2021. 1, 2

    2021

  11. [11]

    Model fusion for building type classification from aerial and street view images.Remote Sensing, 11(11):1259, 2019

    Eike Jens Hoffmann, Yuanyuan Wang, Martin Werner, Jian Kang, and Xiao Xiang Zhu. Model fusion for building type classification from aerial and street view images.Remote Sensing, 11(11):1259, 2019. 1, 2

    2019

  12. [12]

    Com- prehensive urban space representation with varying numbers of street-level images.Computers, Environment and Urban Systems, 106:102043, 2023

    Yingjing Huang, Fan Zhang, Yong Gao, Wei Tu, Fabio Duarte, Carlo Ratti, Diansheng Guo, and Yu Liu. Com- prehensive urban space representation with varying numbers of street-level images.Computers, Environment and Urban Systems, 106:102043, 2023. 2 10

    2023

  13. [13]

    Extensive exposure mapping in urban areas through deep analysis of street-level pictures for floor count determination.Urban Science, 1(2):16, 2017

    Gianni Cristian Iannelli and Fabio Dell’Acqua. Extensive exposure mapping in urban areas through deep analysis of street-level pictures for floor count determination.Urban Science, 1(2):16, 2017. 2

    2017

  14. [14]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. InInter- national conference on machine learning, pages 2127–2136. PMLR, 2018. 4

    2018

  15. [15]

    Perceiver IO: A General Architecture for Structured Inputs & Outputs

    Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Kop- pula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs.arXiv preprint arXiv:2107.14795, 2021. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  16. [16]

    Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set.IEEE Transactions on geoscience and remote sensing, 57(1):574–586, 2018. 2

    2018

  17. [17]

    A scale robust con- volutional neural network for automatic building extraction from aerial and satellite imagery.International journal of remote sensing, 40(9):3308–3322, 2019

    Shunping Ji, Shiqing Wei, and Meng Lu. A scale robust con- volutional neural network for automatic building extraction from aerial and satellite imagery.International journal of remote sensing, 40(9):3308–3322, 2019. 2

    2019

  18. [18]

    Navjot Kaur, Cheng-Chun Lee, Ali Mostafavi, and Ali Mahdavi-Amiri. Large-scale building damage assessment using a novel hierarchical transformer architecture on satel- lite images.Computer-Aided Civil and Infrastructure Engi- neering, 38(15):2072–2091, 2023. 2

    2072

  19. [19]

    Cnn algorithm for roof detection and material classi- fication in satellite images.Electronics, 10(13):1592, 2021

    Jonguk Kim, Hyansu Bae, Hyunwoo Kang, and Suk Gyu Lee. Cnn algorithm for roof detection and material classi- fication in satellite images.Electronics, 10(13):1592, 2021. 2

    2021

  20. [20]

    Enhanced facade parsing for street-level images using convolutional neural networks

    Gefei Kong and Hongchao Fan. Enhanced facade parsing for street-level images using convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 59 (12):10519–10531, 2020. 2

    2020

  21. [21]

    Dominik Laupheimer, Patrick Tutzauer, Norbert Haala, and Marc Spicker. Neural networks for the classification of build- ing use from street-view imagery.ISPRS Annals of the Pho- togrammetry, Remote Sensing and Spatial Information Sci- ences, 4:177–184, 2018. 2

    2018

  22. [22]

    Take a look around: using street view and satellite images to estimate house prices.ACM Transactions on Intelligent Systems and Technology (TIST), 10(5):1–19, 2019

    Stephen Law, Brooks Paige, and Chris Russell. Take a look around: using street view and satellite images to estimate house prices.ACM Transactions on Intelligent Systems and Technology (TIST), 10(5):1–19, 2019. 1, 2

    2019

  23. [23]

    Semi- supervised learning from street-view images and open- streetmap for automatic building height estimation.arXiv preprint arXiv:2307.02574, 2023

    Hao Li, Zhendong Yuan, Gabriel Dax, Gefei Kong, Hongchao Fan, Alexander Zipf, and Martin Werner. Semi- supervised learning from street-view images and open- streetmap for automatic building height estimation.arXiv preprint arXiv:2307.02574, 2023. 2

    work page arXiv 2023

  24. [24]

    Vision foundation models in remote sensing: A survey.IEEE Geoscience and Remote Sensing Magazine, 2025

    Siqi Lu, Junlin Guo, James R Zimmer-Dauphinee, Jordan M Nieusma, Xiao Wang, Steven A Wernke, Yuankai Huo, et al. Vision foundation models in remote sensing: A survey.IEEE Geoscience and Remote Sensing Magazine, 2025. 2

    2025

  25. [25]

    Planet dump re- trieved from https://planet.osm.org .https : //www.openstreetmap.org, 2025

    OpenStreetMap contributors. Planet dump re- trieved from https://planet.osm.org .https : //www.openstreetmap.org, 2025. 3

    2025

  26. [26]

    DINOv2: Learning Robust Visual Features without Supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Using google street view to audit neighborhood environ- ments.American journal of preventive medicine, 40(1):94– 100, 2011

    Andrew G Rundle, Michael DM Bader, Catherine A Richards, Kathryn M Neckerman, and Julien O Teitler. Using google street view to audit neighborhood environ- ments.American journal of preventive medicine, 40(1):94– 100, 2011. 1

    2011

  28. [28]

    Self-supervised vision transformers for land- cover segmentation and classification

    Linus Scheibenreif, Jo ¨elle Hanna, Michael Mommert, and Damian Borth. Self-supervised vision transformers for land- cover segmentation and classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1422–1431, 2022. 2

    2022

  29. [29]

    Understanding urban landuse from the above and ground perspectives: A deep learning, multimodal solution.Remote sensing of environment, 228:129–143, 2019

    Shivangi Srivastava, John E Vargas-Munoz, and Devis Tuia. Understanding urban landuse from the above and ground perspectives: A deep learning, multimodal solution.Remote sensing of environment, 228:129–143, 2019. 2

    2019

  30. [30]

    Esra Suel, Samir Bhatt, Michael Brauer, Seth Flaxman, and Majid Ezzati. Multimodal deep learning from satellite and street-level imagery for measuring income, overcrowding, and environmental deprivation in urban areas.Remote Sens- ing of Environment, 257:112339, 2021. 2

    2021

  31. [31]

    Hierar- chynet: Hierarchical cnn-based urban building classification

    Salma Taoufiq, Bal ´azs Nagy, and Csaba Benedek. Hierar- chynet: Hierarchical cnn-based urban building classification. Remote Sensing, 12(22):3794, 2020. 2

    2020

  32. [32]

    Improving facade parsing with vi- sion transformers and line integration.Advanced Engineer- ing Informatics, 60:102463, 2024

    Bowen Wang, Jiaxin Zhang, Ran Zhang, Yunqin Li, Liangzhi Li, and Yuta Nakashima. Improving facade parsing with vi- sion transformers and line integration.Advanced Engineer- ing Informatics, 60:102463, 2024. 2

    2024

  33. [33]

    Building extraction with vision transformer.IEEE Transac- tions on Geoscience and Remote Sensing, 60:1–11, 2022

    Libo Wang, Shenghui Fang, Xiaoliang Meng, and Rui Li. Building extraction with vision transformer.IEEE Transac- tions on Geoscience and Remote Sensing, 60:1–11, 2022. 2

    2022

  34. [34]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021. 3

    2021

  35. [35]

    Flood vulnerability assess- ment of urban buildings based on integrating high-resolution remote sensing and street view images.Sustainable Cities and Society, 92:104467, 2023

    Ziyao Xing, Shuai Yang, Xuli Zan, Xinrui Dong, Yu Yao, Zhe Liu, and Xiaodong Zhang. Flood vulnerability assess- ment of urban buildings based on integrating high-resolution remote sensing and street view images.Sustainable Cities and Society, 92:104467, 2023. 2

    2023

  36. [36]

    Building damage detection in satellite im- agery using convolutional neural networks.arXiv preprint arXiv:1910.06444, 2019

    Joseph Z Xu, Wenhan Lu, Zebo Li, Pranav Khaitan, and Va- leriya Zaytseva. Building damage detection in satellite im- agery using convolutional neural networks.arXiv preprint arXiv:1910.06444, 2019. 2

    work page arXiv 1910

  37. [37]

    Deep cnn-based methods to evalu- ate neighborhood-scale urban valuation through street scenes perception

    Junhan Zhao, Xiang Liu, Yanqun Kuang, Yingjie Victor Chen, and Baijian Yang. Deep cnn-based methods to evalu- ate neighborhood-scale urban valuation through street scenes perception. In2018 IEEE third international conference on data science in cyberspace (dsc), pages 20–27. IEEE, 2018. 2

    2018

  38. [38]

    Building extraction from satellite images using mask r-cnn with building boundary regularization

    Kang Zhao, Jungwon Kang, Jaewook Jung, and Gunho Sohn. Building extraction from satellite images using mask r-cnn with building boundary regularization. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 247–251, 2018. 2 11

    2018