Seminar: Using synthetic data and domain adaptation for detecting marine species, 21st Nov, 1:00pm

When: Thursday 21st of Nov, 1:00pm 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

Title: Using synthetic data and domain adaptation for detecting marine species

Abstract:

Photographic surveys by autonomous underwater vehicles (AUVs) and other underwater platforms provide a valuable method for monitoring the benthic environment.  Scientists can identify the presence and abundance of benthic species by manually annotating the image using online software or other tools.  Neural network object detectors can reduce the effort involved in this process by locating and classifying the species of interest in the image but accurate detectors often rely on large amounts of annotated training images not currently available.  Our work uses a pipeline to generate large amounts of synthetic annotated training data for a species of interest using 3D imaging software.  Training the detector with the synthetic data can train a detector which can be further improved by adapting the detector model using real unlabelled images.  Our pipeline has produced a detector of sea urchins with comparable performance to a detector trained with manually labelled real images.  Using realistic synthetic data for rare species is a promising approach to reducing the manual effort to analyse imaging survey data.

Bio:

Heather Doig is a PhD student with the Australian Centre for Robotics. She is working on improving automated observations of the underwater environment by applying current machine learning techniques to underwater images. Heather has a Bachelor of Engineering – Mechanical and Manufacturing (Honours) from the University of Melbourne and more recently a Masters of Data Science and Executive Masters in Arts and Social Sciences from the University of Sydney.