Omnidirectional surveys

SeaOrb uses self-stabilizing consumer 360-degree cameras to record the environment surrounding each observation. A complete system costs about $1,000, performs in low light, and requires less specialized framing than a conventional underwater camera.

Learning from marine video

We are training Ocean Mind, a computer-vision model that learns from open marine footage and a curated collection of underwater 360-degree video. Self-supervised learning allows it to learn broadly useful visual features before being trained for specific ecological classifications.

From classifications to ecological maps

Experts first annotate selected points as vegetation, macrofauna, or substrate. A center-pixel classifier then applies those classifications throughout the spherical video.

These predictions provide training data for segmentation models that classify every pixel. The aim is to turn video into spatially explicit measurements of habitat composition, vegetation, and animals.

Focused annotation

Our web platform identifies informative and unusual examples within contributed footage. This directs expert attention toward annotations that add the most value instead of requiring every frame to be reviewed manually.

A reusable foundation

SeaOrb is being developed as an adaptable monitoring project rather than a single-purpose model. Models, annotations, and datasets created for one application can support future work across other species, habitats, regions, and research questions.