SQuID

Gold definitionUpdated Apr 2, 2026

Definition

SQuID (Satellite Quantitative Intelligence Dataset) is a benchmark of 2,000 satellite image Question-Answer pairs designed to evaluate quantitative spatial reasoning in Vision-Language Models. It features numerical range and categorical answers across three difficulty tiers.

At a glance

Executive summary

SQuID is a new dataset that helps test how well AI models can understand and answer questions about satellite images, especially those involving counting or measuring things. It was created because current AI models aren't very good at these precise visual tasks, which require detailed pixel information.

TL;DR

SQuID is a new dataset of satellite image questions and answers designed to challenge AI models to get better at precise counting and measuring in pictures.

Key points

A benchmark dataset with satellite image Q&A pairs for quantitative spatial reasoning.
Evaluates and highlights limitations of Vision-Language Models in precise pixel-level spatial reasoning (counting, measurements).
Used by researchers and engineers developing Vision-Language Models and spatial AI.
Unlike general VLM benchmarks, SQuID specifically targets quantitative spatial reasoning, which is often overlooked or poorly handled by existing benchmarks.
Drives development of Vision-Language Models with improved pixel-level precision and code-generation architectures for visual analysis.

Use cases

Evaluating new Vision-Language Models for their ability to accurately count objects in aerial imagery (e.g., number of cars, buildings).
Benchmarking AI systems designed for environmental monitoring, such as measuring deforestation or urban expansion from satellite data.
Developing AI for disaster response, where precise quantification of damaged areas or deployed assets from aerial views is critical.
Training and validating models for agricultural analysis, like estimating crop yield or identifying specific plant counts from drone or satellite images.

Also known as

Satellite Quantitative Intelligence Dataset