RarePlanes: Real and Synthetic Satellite Imagery for Aircraft Detection

The RarePlanes dataset incorporates both real and synthetically generated satellite imagery to test the value of synthetic data for computer vision. It consists of 253 real Maxar WorldView-3 satellite scenes with 14,700 hand-annotated aircraft and 50,000 synthetic images with approximately 630,000 aircraft annotations. The dataset was created by CosmiQ Works and AI.Reverie and is hosted on AWS Open Data.

Use Cases

• Training aircraft detection models based on the 14,700 hand-annotated real satellite images
• Evaluating the utility of synthetic data for model generalization based on the 50,000 synthetic images
• Developing algorithms to classify aircraft attributes based on the provided annotations
• Benchmarking object detection performance on very high resolution overhead imagery

Strengths

• Includes 14,700 hand-annotated aircraft in real satellite imagery, providing a high-quality validation set
• Contains a large synthetic component of 50,000 images with ~630,000 annotations for augmentation studies
• Real data covers 2,142 km^2 across 112 locations, offering geographic diversity

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Last update date is unknown; freshness unverified

Provenance

Source

CosmiQ Works and AI.Reverie

Collection Method

Real data from Maxar WorldView-3 satellite scenes; synthetic data generated via AI.Reverie's simulation platform.

Time Range

null

Freshness

Last updated is unknown.

Geography

Spans 112 locations (specific locations not provided).