Recent Research
Embedded Files
Street-Level Geolocation from Online Videos: A Framework for Cross-View Extraction and Map Alignment (2025, Ongoing)
Abstract:
Street-level videos available online offer a vast, untapped source of geospatial information. In this work, we introduce a framework that automatically extracts high-quality street view images from online videos and accurately geolocates them on a map, even when no GPS data is available. Our method addresses three main challenges: identifying useful street-view frames, estimating their geographic location through cross-view matching, and aligning sequences of frames with real-world map layers.
Hierarchical Change Detection for Street-Level Imagery Using Satellite Data (2025, Ongoing)
Abstract:
Street-level environments change rapidly, but capturing these changes at scale remains challenging. In this work, we develop a hierarchical monitoring framework that leverages low-resolution satellite imagery to predict areas of potential street-level change, enabling selective and efficient targeting of high-resolution ground imagery collection. By combining broad, low-cost satellite coverage with precise street-level updates, our approach accelerates the detection of urban development, infrastructure changes, and environmental shifts. This framework optimizes resource allocation for ground data collection, supporting scalable, near-real-time street-level monitoring across diverse regions
Automated Retrieval and Pixel-Accurate Video Generation from Satellite Imagery (2024–2025, Ongoing)
Abstract:
Satellite imagery archives contain immense temporal and spatial information, but retrieving and comparing changes at pixel-level accuracy remains a major challenge. In this work, we develop a framework that automatically retrieves time-separated satellite images over specified regions and generates comparison videos with pixel-accurate registration at scale. Our approach addresses three key problems: identifying and retrieving relevant satellite scenes, aligning imagery to sub-pixel precision across time periods, and generating smooth, interpretable visualizations that highlight real-world changes for analysis and modeling.
Cross-Domain Image Upscaling for Satellite and Street-Level Imagery (2024)
Abstract:
High-resolution imagery is essential for detailed geospatial analysis, but satellite and street-level images often suffer from sensor limitations, noise, and scale variability. In this work, we explore cross-domain super-resolution by developing a mixture of specialized models rather than relying on a single architecture. Our framework combines domain-adapted upscaling models to address the unique challenges of satellite and street-level imagery, including preserving structural consistency, minimizing modality-specific artifacts, and maintaining semantic integrity. By orchestrating multiple expert models, we aim to produce sharper, more reliable imagery that supports improved feature recognition, change detection, and AI model training across diverse environments.
Training Visual-Geospatial Reasoning through Satellite Image Sequencing (2023)
Abstract:
Understanding temporal and spatial patterns in satellite imagery is critical for geospatial analysis, yet developing intuitive reasoning skills remains a challenge. In this work, we design a training framework that presents users with shuffled sequences of satellite images and tasks them with reconstructing the correct temporal order and inferring geographic location. This system supports the development of visual-geospatial intuition, change detection skills, and temporal reasoning. Beyond training applications, the framework enables the generation of self-supervised learning datasets, supports disaster response simulations, and offers new pathways for AI model pretraining focused on environmental understanding and sequence prediction.
Page updated
Report abuse