Separate Registration Required

Imagine being tasked with finding instances of a rarely seen object across a massive archive of overhead imagery—without an AI model trained for it. Traditionally, this would require curating and labeling hundreds or thousands of examples, followed by model training—a process that’s costly, time-consuming, and resource-intensive.

Now, emerging Vision-Language Models (VLMs) are changing that paradigm. By allowing users to simply describe what they’re looking for in natural language, VLMs can identify and segment objects without any new annotations or retraining. Trained on internet-scale datasets of images and text, these foundation models learn a shared feature space that connects visual and linguistic information. Though not purpose-built for geospatial data, they can be adapted effectively for tasks such as object detection and semantic segmentation in overhead imagery.

This session demystifies VLMs for a geospatial audience, focusing on their practical applications in object detection, segmentation, and 3D analysis. Attendees will learn how these models work conceptually, how to use them to build detectors from natural language prompts, and how to extend their use to pixel-wise and 3D segmentation in meshes and point clouds.

Key Topics

How VLMs encode and align visual and textual information
Detecting objects with no prior training data (zero-shot learning)
Extending VLMs for semantic segmentation of imagery
Applying image-based VLMs to 3D geospatial data
Understanding current limitations and mitigation strategies
Adapting VLMs to new concepts with limited image examples (few-shot learning)

Learning Outcomes

Participants will gain a clear understanding of how vision-language models can be applied to object detection and semantic segmentation in geospatial contexts. The session will help attendees evaluate whether VLMs are appropriate for their mission or analytical needs and understand when they offer an alternative to traditional annotation-heavy model development.

Prerequisites

None required. This session is designed for a broad audience—including engineers, analysts, decision-makers, and policy professionals—interested in object detection and segmentation within geospatial imagery and 3D data. While technical concepts will be discussed, explanations will remain high-level and accessible. No coding is required; references to open-source tools and models will be provided for those who wish to explore further.

GEOINT 2026

Detecting and Segmenting Anything You Can Put into Words with VLMs

Type

Speakers

Description

Tracks