Seminar: Reducing Label Dependency for Automated Analysis of Underwater Imagery, 5th Dec, 10:00am

When: Thursday 5th of Dec, 10:00am AEDT

Where: This seminar will be partially presented at the ACFR seminar area, J04 lvl 2 (Rose St Building) and partially online via Zoom. RSVP

Speaker: Dr Scarlett Raine

Title: Reducing Label Dependency for Automated Analysis of Underwater Imagery

Abstract:

This seminar presents novel deep learning approaches which reduce label dependency for automated analysis of imagery collected by robotic underwater and surface vehicles. With the increasing use of robotics to study coral reefs and seagrass meadows, vast amounts of imagery are generated. Traditionally, analysing this data has been challenging, time-consuming, and expensive, as it heavily relied on marine experts. This seminar describes innovative artificial intelligence methods which automatically identify marine species and reduce the annotation effort for domain experts. These contributions span design of data collection and annotation methodology; seagrass segmentation from image-level labels only; and use of large language models as a supervisory signal in domain-specific applications. This seminar will also describe a human-in-the-loop labelling regime in which the model and a domain expert work together to improve annotation efficiency for coral images. Finally, the seminar will discuss our current research with the Reef Restoration and Adaptation Program and the Australian Institute of Marine Science on the Deployment Guidance System (DGS). The DGS aims to guide the automated deployment of coral re-seeding devices to areas of the seafloor that provide optimal survival conditions.

Bio:

Dr Scarlett Raine is a Research Fellow in the QUT Centre for Robotics. She brings her expertise in artificial intelligence to the Reef Restoration and Adaptation Program, where she is working on the Transition to Deployment sub-program.  She recently completed her PhD on the topic of Weakly Supervised Segmentation of Underwater Imagery, which was a collaboration between the QUT Centre for Robotics and CSIRO’s Data61. Scarlett is a published researcher in computer vision and artificial intelligence for challenging applied robotics and environmental monitoring problems, with papers published in the Robotics and Automation Letters, WACV, CVPR-W, IROS and DICTA. Scarlett is motivated by data-constrained and weakly-labelled problems, and automated analysis of real-world field data, with a particular focus on conservation of coral reefs, seagrass meadows and coastal ecosystems.

Contacts

Australian Centre for Robotics
info@acfr.usyd.edu.au