Article
Computer Science, Artificial Intelligence
Yun Liu, Ming-Ming Cheng, Deng-Ping Fan, Le Zhang, Jia-Wang Bian, Dacheng Tao
Summary: A novel fully convolutional neural network is proposed for semantic edge detection, using diverse deep supervision within a multi-task framework to overcome the challenge of distinct supervision targets. The network aims to locate fine detailed edges and identify high-level semantics simultaneously.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2022)
Article
Computer Science, Artificial Intelligence
Xiaolin Hu, Chufeng Tang, Hang Chen, Xiao Li, Jianmin Li, Zhaoxiang Zhang
Summary: This article introduces a post-processing refinement framework called BPR to improve the boundary quality of predicted image segmentation models. By extracting and refining boundary patches, the proposed method addresses the issue of imprecise boundaries and achieves significant improvements.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2022)
Article
Engineering, Electrical & Electronic
Yusen Xie, Shaolan Zheng, Haiyun Wang, Yuzhou Qiu, Xilan Lin, Qian Shi
Summary: Farmland is a crucial resource for human survival and development. Rapid acquisition of farmland information is important for crop detection and land development sustainability. High-resolution remote sensing imagery and deep learning-based image segmentation methods have been widely used in remote sensing, but have difficulties in extracting refined farmland parcels. We propose a method that utilizes deep neural networks for farmland edge detection, high-resolution networks for feature extraction, and a postprocessing method for connecting disconnected boundaries. Experimental verification using Google Earth images shows higher precision and more detailed and complete farmland parcels.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2023)
Article
Geochemistry & Geophysics
Konrad Heidler, Lichao Mou, Celia Baumhoer, Andreas Dietz, Xiao Xiang Zhu
Summary: This study proposes a new model that combines deep learning methods for segmenting and delineating coastlines. By combining building blocks from different frameworks and using deep supervision and hierarchical attention mechanism, the training effectiveness is improved. The advantages of this approach over traditional methods and other deep learning methods are demonstrated on a challenging dataset of Antarctic coastlines.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Engineering, Electrical & Electronic
Iaroslav Koshelev, Maxim Savinov, Alexander Menshchikov, Andrey Somov
Summary: In this article, the problem of hogweed detection using a drone equipped with RGB and multispectral cameras is addressed. Two approaches are studied: offline detection running on the orthophoto and real-time scanning from the frame stream on the edge device. The introduction of an additional convolution neural network trained with transfer learning helps boost the detection quality and eliminate the need for a multispectral camera.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2023)
Article
Remote Sensing
Maxwell Jong, Kaiyu Guan, Sibo Wang, Yizhi Huang, Bin Peng
Summary: Field boundary data is essential for digital agricultural services and research, and adversarial training can significantly improve the quality and performance of field boundary prediction.
INTERNATIONAL JOURNAL OF APPLIED EARTH OBSERVATION AND GEOINFORMATION
(2022)
Article
Physics, Multidisciplinary
Lintao Yu, Anni Yao, Jin Duan
Summary: In this paper, a method that utilizes decoupling and combines edge information for semantic segmentation is proposed. A new dual-stream CNN architecture is built to fully consider the interaction between the body and the edge of the object, resulting in significant improvement in the segmentation performance of small objects and object boundaries.
Article
Computer Science, Information Systems
Yi Li, Zhangbing Zhou, Xiao Xue, Deng Zhao, Patrick C. K. Hung
Summary: This article proposes an accurate anomaly detection mechanism with energy efficiency in three-tier IoT-edge-cloud collaborative networks. It filters anomaly-relevant sensory data at the edge tier to decrease network traffic. The boundary of anomaly is determined using the Kriging spatial interpolation algorithm at the cloud tier and refined using mobile sensing nodes at edge networks.
IEEE INTERNET OF THINGS JOURNAL
(2023)
Article
Geochemistry & Geophysics
Jianhui Jin, Wujie Zhou, Rongwang Yang, Lv Ye, Lu Yu
Summary: The acquisition of high-resolution satellite and airborne remote sensing images has become easier due to the development of sensor technology. Semantic segmentation based on multi-source information fusion is gaining popularity as single-modal images struggle to accurately classify objects in complex scenes. A proposed multimodal fusion network addresses this issue by utilizing edge detection and boundary information.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2023)
Article
Computer Science, Information Systems
Ming-Gang Gan, Yan Zhang
Summary: This paper proposes a framework called DHCNet for temporal action detection. The framework achieves accurate classification of action proposals through the structure of two subnets and a coarse-to-fine classification method, and improves the temporal boundaries through an iterative boundary regression method. Experiments demonstrate the effectiveness of this approach.
MULTIMEDIA TOOLS AND APPLICATIONS
(2023)
Article
Mathematics
Yongxu Liu, Zhi Zhang, Yan Liu, Yao Zhu
Summary: Non-invasive neuroimaging techniques and graph theories have played a crucial role in understanding the structural patterns of the human brain. However, research on bad-channel detection, a task of imbalanced classification, is limited. In this study, a novel edge generator is proposed, taking into account the prominent small-world organization of the human brain network.
Article
Computer Science, Artificial Intelligence
Dongyue Wu, Zilin Guo, Aoyan Li, Changqian Yu, Changxin Gao, Nong Sang
Summary: In this paper, a novel conditional boundary loss (CBL) is proposed to improve the boundary segmentation results in semantic segmentation. The CBL optimizes each boundary pixel by pulling it closer to its local class center and pushing it away from different-class neighbors. Experimental results demonstrate that applying the CBL to popular segmentation networks significantly improves the mIoU and boundary F-score performance.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2023)
Article
Biochemistry & Molecular Biology
Taukir Alam, Wei-Chung Shia, Fang-Rong Hsu, Taimoor Hassan
Summary: This research analyzes and evaluates breast cancer detection and diagnosis using segmentation models. The Unet3+ model is found to have optimal performance, with an average accuracy of 82.53% and an average intersection over union (IU) of 52.57%. The application of these models shows remarkable results and has the potential to improve patient outcomes.
Article
Computer Science, Information Systems
Juuso Eronen, Michal Ptaszynski, Fumito Masui, Aleksander Smywinski-Pohl, Gniewosz Leliwa, Michal Wroczynski
Summary: The study investigates the effectiveness of Feature Density (FD) with linguistically-backed feature preprocessing methods to estimate dataset complexity, aiming to comparatively estimate the potential performance of ML classifiers and reduce the number of required experiments iterations. This is crucial for optimizing resource-intensive training of ML models, especially in the context of increasing dataset sizes and the environmental impact of CO2 emissions from large-scale model training.
INFORMATION PROCESSING & MANAGEMENT
(2021)
Article
Computer Science, Artificial Intelligence
Linh T. Duong, Cong Q. Chu, Phuong T. Nguyen, Son T. Nguyen, Binh Q. Tran
Summary: This study proposes a practical solution for classifying mammograms using the synergy between graph neural networks and image processing techniques. The experimental results demonstrate that the proposed approach achieves optimal prediction performance on the dataset, achieving 100% accuracy and 1.0 precision and recall in the classification of BI-RADS scores and breast density types. The proposed approach is anticipated to be deployed as a non-invasive pre-screening tool to assist doctors in their diagnosis activities.
APPLIED SOFT COMPUTING
(2023)
Article
Construction & Building Technology
Rong Huang, Yusheng Xu, Ludwig Hoegner, Uwe Stilla
Summary: This study proposes a semantic-aided change detection method to monitor construction progress using UAV-based photogrammetric point clouds. The framework consists of two key parts that identify geometric and semantic changes progressively. By considering different types of changes, the method can fully consider the changes that may occur at construction sites.
AUTOMATION IN CONSTRUCTION
(2022)
Article
Geochemistry & Geophysics
Rong Huang, Wei Yao, Yusheng Xu, Zhen Ye, Uwe Stilla
Summary: In this study, a coarse-to-fine registration strategy utilizing rotation-invariant features and a graph matching (GM) method was developed for point cloud processing. The method successfully achieved accurate correspondence search, with experimental results showing fine registration capabilities.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2022)
Article
Construction & Building Technology
Theresa Meyer, Ansgar Brunn, Uwe Stilla
Summary: This paper presents a method for high-resolution change detection in confined building interiors using terrestrial laser scanning and voxel-based spatial discretization. The method allows automatic construction progress documentation and metric evaluation. By modeling the effects of laser range measurements on the occupancy of space and evaluating using evidence theory, the method can verify, update, and confirm the metric accuracy of a BIM.
AUTOMATION IN CONSTRUCTION
(2022)
Article
Geography, Physical
Theresa Meyer, Ansgar Brunn, Uwe Stilla
Summary: This paper presents a method for the geometric verification of indoor BIMs by images and uncertainty management to exploit the potential of photogrammetry in the context of professional building documentation and digital twinning. Individual 3D point accuracies, object surface characteristics, and BIM related uncertainties are assessed and considered. A combined reasoning pipeline based on Dempster-Shafer evidence theory is used to make the final decision on whether a photogrammetric point cloud meets the required level of accuracy for a given model. The novel Pho-to-BIM verification method is demonstrated on three real indoor construction sites, showing how belief functions can be set up individually based on measurement and site characteristics.
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
(2023)
Article
Engineering, Civil
Yan Xia, Qiangqiang Wu, Wei Li, Antoni B. B. Chan, Uwe Stilla
Summary: This paper proposes DMT, a Detector-free Motion-prediction-based 3D Tracking network, which leverages temporal motion cues to address the target-specific detection problem in sparse and incomplete LiDAR point clouds. Experimental results demonstrate that DMT is lighter, faster, and more accurate than previous trackers, achieving better performance (around 10% improvement over the NuScenes dataset) without the need for complicated 3D detectors.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2023)
Review
Geography, Physical
Uwe Stilla, Yusheng Xu
Summary: This article provides a comprehensive review of point-cloud-based 3D change detection for urban objects. The study aims to identify critical techniques and explore the applications of point clouds in various fields such as land cover monitoring and transportation monitoring. The limitations of current change detection technology and research gaps are also discussed.
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
(2023)
Article
Geography, Physical
Xinyi Li, Yinlong Liu, Yan Xia, Venkatnarayanan Lakshminarasimhan, Hu Cao, Feihu Zhang, Uwe Stilla, Alois Knoll
Summary: This research focuses on 4 degrees of freedom point set registration and proposes a method to accelerate branch and bound-based methods by decoupling the joint pose. Experimental results show that the proposed method is more robust and faster than existing methods, and can solve the challenging problem of simultaneous pose and correspondence registration.
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
(2023)
Article
Environmental Sciences
Manoj Kumar Biswanath, Ludwig Hoegner, Uwe Stilla
Summary: This paper proposes a framework for mapping temperature attributes from thermal point clouds onto building facades, aiming to generate thermal textures for 3D analysis. The approach involves using point clouds from mobile laser scanning and intensities extracted from thermal infrared image sequences. A mapping algorithm based on nearest neighbor search is used to project the thermal point clouds onto facades. The generated texture is evaluated based on a performance metric called root-mean-square deviation (RMSD), with the nearest neighbor method outperforming other interpolation methods.
Article
Geochemistry & Geophysics
Shuo Shen, Yan Xia, Andreas Eich, Yusheng Xu, Bisheng Yang, Uwe Stilla
Summary: Three-dimensional point cloud semantic segmentation is crucial for scene understanding, but existing solutions struggle to generalize well to new data with different sensor configurations. To address this problem, we propose SegTrans, an unsupervised domain adaptation method that greatly improves the generalization performance from a labeled dataset (source domain) to an unlabeled dataset (target domain) for point cloud semantic segmentation.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2023)
Article
Engineering, Electrical & Electronic
Jingwei Zhu, Yusheng Xu, Ludwig Hoegner, Uwe Stilla
Summary: Monitoring building efficiency is a hot topic, and thermal infrared (TIR) images are commonly used. However, professional knowledge is required for analysis, and images cannot fully describe all aspects of the thermal attributes. As a solution, as-built thermal point clouds can be generated using mobile laser scanning (MLS) and provide better calibration results than manual methods.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2023)
Proceedings Paper
Geography, Physical
O. Wysocki, E. Grilli, L. Hoegner, U. Stilla
Summary: In this paper, we propose a method for enriching 3D building models with facade openings by combining visibility analysis and neural networks. The method uses occupancy voxels to provide semantics and detect conflicts, and combines the results in a Bayesian network to classify and delineate facade openings. The method has been tested on various datasets and has been proven to improve semantic segmentation accuracy and upgrade building models.
17TH 3D GEOINFO CONFERENCE
(2022)
Proceedings Paper
Geography, Physical
Lukas Lucks, Philipp-Roman Hirt, Ludwig Hoegner, Uwe Stilla
Summary: This paper presents a method for monitoring Alpine slope movements based on image sequences. By calculating the object coordinates of key points and detecting correspondences, three-dimensional motion vectors can be determined. The analysis of landslide image sequences shows an average movement of 75 mm.
XXIV ISPRS CONGRESS IMAGING TODAY, FORESEEING TOMORROW, COMMISSION II
(2022)
Article
Engineering, Electrical & Electronic
Carlos Villamil Lopez, Uwe Stilla
Summary: In this article, a novel SAR change detection method is proposed for the monitoring of man-made objects. The method detects changes by identifying the appearance and disappearance of strong scatterers, ignoring changes to natural targets. Additionally, an object-based change analysis step is introduced.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2022)
Article
Engineering, Mechanical
Christian Rudolf Albrecht, Jenny Behre, Eva Herrmann, Stefan Juergens, Uwe Stilla
Summary: This study introduces a novel robustness score that combines different aspects of robustness and evaluates a graph-based localization method with fault injections. It also explores the impact of semantic class information on localization system robustness and how it can improve false-positive rates.
Article
Remote Sensing
Olaf Wysocki, Ludwig Hoegner, Uwe Stilla
Summary: This paper presents a method of reconstructing underpasses by comparing building models' facades with co-registered MLS measurements, aiming to enhance semantic 3D building models. Experimental results show an accuracy of 12 cm with differences in volumes (up to 18%) and surfaces (up to 20%) between reconstructed and updated models.
INTERNATIONAL JOURNAL OF APPLIED EARTH OBSERVATION AND GEOINFORMATION
(2022)