cranfield-synthetic-drone-detection

A fully synthetic dataset of drones generated using structured domain randomization. It contains multiple datasets generated using different styles:

• Drones only
• Drones and Birds
• Generic Distractors
• Realistic Distractors
• Random Backgrounds

This was done to understand how different styles affect the sim-to-real translation of faster-rcnn models (full study is done in the paper)

Further, the bounds of the camera in relation to the drones is a simulation parameter that has a strong influence on the size of the generated drones in the dataset. To understand this relation, multiple datasets are generated, with different maximum camera bounds:

• 20 m
• 40 m
• 80 m
• 160 m
• 320 m A study on how this affects transferability to real life datasets is conducted in the paper.