MVRL is a computer vision research lab developing multimodal representation learning, geospatial AI, generative modeling, and related methods at planetary scale. We validate and deploy these methods across application areas spanning biodiversity and conservation, agriculture, urban sensing, geohazards, air quality, transportation safety, and medical imaging.
The lab’s core technical agenda. We develop self-supervised and multimodal representation learning, planet-scale geospatial AI, generative modeling for images and 3D scenes, geometric vision and localization, vision-language models, reinforcement learning for active search, and uncertainty estimation for safety-sensitive deployments.
Recent activity: 15 papers (2024–2026)
How do we learn representations that transfer across sensors, modalities, and domains? We develop self-supervised and multimodal embedding methods that align vision, language, and audio for robust learning at scale....
Recent activity: 49 papers (2024–2026)
How do we build geospatial machine learning that works reliably across views, resolutions, and sensors worldwide? Geospatial AI is our platform for planet-scale understanding: unified embeddings, cross-view localization, retrieval, and...
Recent activity: 20 papers (2024–2026)
How can generative models synthesize and represent information across sensors, scales, and modalities? We develop generative methods for images, 3D scenes, panoramas, and multimodal earth data. Recent work includes open-world...
Recent activity: 6 papers (2024–2026)
How do we recover geometry, depth, and 3D structure from images and panoramas? Geometric vision supports scene understanding, generative modeling, and geospatial applications across the lab. Recent work includes open-world...
Recent activity: 8 papers (2024–2026)
How can vision-language models retrieve, compose, and reason over images with text? We develop and apply vision-language methods for retrieval, captioning, and multimodal understanding in geospatial and natural-world settings. Recent...
Recent activity: 7 papers (2024–2026)
How should an agent search visual or geospatial space when observations are partial and costly? Our reinforcement learning research targets active search, exploration, and sequential decision-making in visual domains. Recent...
Recent activity: 3 papers (2024–2026)
Where on Earth was an image taken—and how do we align ground-level views with overhead reference data? Image localization is a long-standing focus of the lab, spanning static outdoor cameras,...
How can we estimate camera pose and intrinsics from images when metadata are missing or unreliable? We develop calibration and pose-estimation methods that exploit structure, natural cues, and cross-view consistency....
When should a model abstain or flag low confidence—and how do we calibrate neural predictors for deployment? We study uncertainty estimation and calibration for vision models, with emphasis on reliable...
Problem domains where we deploy and validate our methods. Activity levels reflect recent publications.
Recent activity: 8 papers (2024–2026)
What is changing in species, habitats, and ecosystems—and how can AI support conservation decisions? We build multimodal models for biodiversity monitoring and environmental change from natural-world imagery, citizen science, and...
Recent activity: 7 papers (2024–2026)
Agricultural applications within MVRL focus on global field boundary mapping and crop monitoring from satellite imagery. Highlights include the Fields of the World (FTW) benchmark and tooling for worldwide agricultural...
Recent activity: 7 papers (2024–2026)
How can overhead imagery, geotagged media, and geospatial AI support understanding of cities and the built environment? We study land use, property and infrastructure patterns, traffic dynamics, and human-centered landscape...
Recent activity: 4 papers (2024–2026)
How can overhead imagery and 3D sensing improve roadway safety assessment and traffic understanding? Transportation research applies computer vision and machine learning to crash risk, traffic dynamics, and infrastructure monitoring....
Recent activity: 3 papers (2024–2026)
How can we map hazardous terrain and environmental disturbances from elevation, LiDAR, and overhead imagery? We develop segmentation and fusion methods for sinkholes, landslides, and post-disturbance landscape change. Recent work...
Recent activity: 2 papers (2024–2026)
How can we monitor forest structure, mortality, and disturbance at scale from satellite time series? We build methods for tree mortality mapping, deforestation detection, and forest typing from Sentinel-2, LiDAR,...
Recent activity: 3 papers (2024–2026)
How can satellite and geophysical data improve estimates of air pollution and atmospheric composition? We develop deep learning methods that fuse remote sensing with meteorological and land-surface variables to map...
How can multimodal representation learning improve diagnostic tools when data are scarce or heterogeneous? We apply visual and multimodal learning methods developed in the lab to medical and biological imaging,...