Reef Scene Understanding
Expert annotation of underwater imagery is slow and expensive. We develop methods that learn from very sparse labels.
Why it matters
Coral reef surveys produce far more images than experts can label. Our point-label-aware superpixels and human-in-the-loop methods turn 5 to 25 labelled points per image into dense segmentation, and image-level labels alone are enough for coarse seagrass mapping. We have also surveyed label-efficient underwater scene understanding and shown how strongly domain and species affect object detection performance.
Team
Members of my group at the QUT Centre for Robotics and close collaborators.

Lead

Collaborator

PhD student
Dorian Tsai
Collaborator
Collaborator
Matthew Dunbabin
Collaborator
Collaborator



