Abstract

Air-sea rescue is a highly time sensitive endeavor. Victim submersion time strongly influences survival time. The problem is exasperated by the vast search spaces presented by the ocean. Finding individuals quickly in such spaces can be challenging. Drones using computer vision can be a great boon in these situations. Computer vision models require vast amounts of labelled data. The acquisition and labelling of said data can make the use of computer vision models infeasible. Synthetic data represents a possible solution. Generated images may present a sufficient facsimile of authentic images to be used in their stead. Our study used the combination of several tools to produce an automated synthetic data pipeline which both generated and labelled images. Our results showed the model performed well even in comparison to studies using purely authentic data.

Download

Share

COinS