The University of Vermont
Vermont Artificial Intelligence Laboratory
Publications
For a full and latest list of publications, please refer to
Cross-view Geo-localization via Learning Disentangled Geometric Layout Correspondence
Our paper "Cross-view Geo-localization via Learning Disentangled Geometric Layout Correspondence" has been accepted to AAAI-23. In this paper, we propose GeoDTR which explicitly disentangles the geometric layout information from low-level details. We also present two categories of augmentations to regularize the model to learn more robust and coherent feature representations. Moreover, a counterfactual training paradigm is adopted to provide an extra weakly supervision signal to the model.
Our work "Cross-View Image Sequence Geo-localization" has been accepted to WACV 2023. In this paper, we propose and tackle the problem of cross-view image sequence geo-localization, which is an extension of the cross-view image geo-localization problem. Since no existing dataset is available, we collected the first large-scale dataset for training and benchmarking purposes for this problem. Moreover, we propose a specific feature aggregation module to exploit the temporal feature in the ground sequence.
For more detailed information, please refer to our paper here. The corresponding code and dataset can be accessed here.
If you are interested in more comprehensive topics in visual geo-localization, our recent papers, "Visual and Object Geo-localization: A Comprehensive Survey" can be a good resource. A preprint version is available here.
Objects detection, geo-localization and enhancement in ground and satellite level images
The concept of geo-localization refers to the process of determining where on earth some ‘entity’ is located, typically using GPS coordinates. The entity of interest may be an image, sequence of images, a video, satellite image, or even objects visible within the image. As massive datasets of GPS tagged media have rapidly become available due to smartphones and the internet, and deep learning has risen to enhance the performance capabilities of machine learning models, the fields of visual and object geo-localization have emerged due to its significant impact on a wide range of applications such as augmented reality, robotics, self-driving vehicles, road maintenance, and 3D reconstruction.
In this research thrust, we investigate advanced deep learning-based approaches to effectively geo-localize images, objects from the ground- and satellite-level images. Which involves either determining from where an image has been captured (Image geo-localization) or geo-locating objects within an image (Object geo-localization).
Almutairy F, Alshaabi T, Nelson J, Wshah S. ARTS: Automotive Repository of Traffic Signs for the United States. IEEE Transactions on Intelligent Transportation Systems. 2019 Dec 27;22(1):457-65.
Wilson D, Alshaabi T, Van Oort C, Zhang X, Nelson J, Wshah S. Object Tracking and Geo-localization from Street Images. arXiv preprint arXiv:2107.06257. 2021 Jul 13.
Wilson D, Zhang X, Sultani W, Wshah S. Visual and Object Geo-localization: A Comprehensive Survey. arXiv preprint arXiv:2112.15202. 2021 Dec 30.
MACHINE LEARNING-GUIDED DESIGN OF ANTIMICROBIAL PEPTIDES
Antimicrobial peptides (AMPs) show promise in treating resistant infections. However, applications of known AMPs have encountered issues in development, production, and shelf-life. To drive the development of AMP-based treatments, it is necessary to create design approaches with higher precision and selectivity toward resistant targets.
In this project, we developed AMPGAN a generative adversarial-based model, and obtained proof-of-concept evidence for the generative approach to design AMPs with experimental validation. We proposed a bidirectional conditional generative adversarial network (BiCGAN)-based approach for rational AMP design. Antimicrobial peptides (AMPs) show promise in treating resistant infections. However, applications of known AMPs have encountered issues in development, production, and shelf-life. To drive the development of AMP-based treatments, it is necessary to create design approaches with higher precision and selectivity toward resistant targets. In this project, we developed AMPGAN a generative adversarial-based model, and obtained proof-of-concept evidence for the generative approach to design AMPs with experimental validation. We proposed a bidirectional conditional generative adversarial network (BiCGAN)-based approach for rational AMP design.
Cyber attacks detection, localization and correction in power systems
Many cyber-attacks, such as false data injection attacks, GPS spoofing attacks, can be undetectable using conventional detection methods. These attacks, if not detected and corrected in real-time, can harm the power grid and might lead to catastrophic consequences. We are proposing advanced deep learning approaches to accurately detect, localize, and correct cyber-attacks in power grids.
For False Data Injection Attacks (FDIAs) we have developed a new methodology for not only identifying AC FDIAs but, more importantly, for correction as well which is we are the first group to correct corrupted data by FDIAs. Our methodology utilizes a Long-Short Term Memory Denoising Autoencoder (LSTM-DAE) to correct attacked-estimated states based on the attacked measurements. Our experiments demonstrated that the proposed method was successfully able to identify the corrupted states and correct them with high accuracy.
In GSAs, the adversary launches interference causing the GPS receiver to lose track, resulting in phase shift measurements propagating not only on the affected PMU but also on the entire network. We proposed novel Transfomerr-BLSTM deep learning method to detect and mitigate the effects of GSAs against PMUs. Our proposed methods are computationally efficient and can accurately detect and mitigate multiple simultaneous GSAs over the entire range of possible phase shifts.