A geospatial vision dataset for detecting, counting, and precisely locating objects smaller than five pixels in satellite and aerial imagery. Spanning wind turbines, offshore platforms, and navigation beacons, it pairs imagery with natural-language questions, object counts, fractional-pixel location annotations, and visual answer keys. Positive examples and challenging negatives test fine-grained perception and false-positive resistance, while source lineage and federal inventory references support traceability.
Available on Request · Sample Available
Sub-5-Pixel Object Detection
Tiny objects. Precise locations. A demanding test of geospatial vision.
Volume
0
Volume
Scalable Volume
Coverage
Multi-Sensor
0.0
%
Targets
Modality & Delivery
imagestext→textjsonimages
Formats
.jsonImage
What You Receive
Coverage
Satellite and aerial imagery covering wind turbines, offshore platforms, and navigation beacons
Sentinel-2 at 10 m/pixel and NAIP at 0.6 m/pixel
Positive, hard-negative, and clean-negative examples
Package Structure
Differentiators
- Fractional-pixel location annotations
- Visual answer keys with magnified object insets
- Per-item imagery provenance and federal inventory references
- Native sensor grids preserved without resampling
Record Schema
Results
Methodology & Quality
Methodology
SME involvement
Stressors
Extremely small visual targets
Shadow-dependent visibility
Confusable structures in negative examples
Different sensor resolutions and ground coverage
Design criteria
- Target objects measure fewer than five pixels
- Every target clears a three-sigma local contrast threshold
- Retain fractional-pixel coordinates
- Include positive and negative examples
Build assumptions
Visibility selection favors sparse, visually confirmable scenes
Quality approach
- Visually inspect annotated answer keys
- Check inventory dates and jurisdiction
- Reject tiles with ambiguous, overlapping, or excluded target annotations
- Record source lineage and visibility metadata