TechSocial Series - December 2024

Abstract

Biologists investigating animal behaviour face a fundamental challenge: biology is inherently noisy, and studying behaviour often requires vast amounts of highly variable data. Off-the-shelf computer vision tools, however, struggle when applied to non-model species, as their training data lack the diversity to generalise across species and environments. To address this, we set out to develop an automated data generation and analysis pipeline comprising scAnt, replicAnt, and OmniTrax, to build robust computer vision models capable of handling real-world challenges (https://evo-biomech.ic.ac.uk/open-science/).

scAnt creates high-resolution 3D scans of small animals, offering detailed morphological data for analysis and forming the basis for our synthetic data generator. replicAnt leverages Unreal Engine 5 to generate diverse, annotated images of animals in complex environments, bridging the gap between controlled lab conditions and unpredictable fieldwork. Finally, OmniTrax enables real-time tracking and pose estimation of large groups of freely moving animals.

By using this automated pipeline, we open up the possibility of studying animals in challenging conditions, with much-reduced costs in manual labour. This approach significantly accelerates the development of robust models, allowing researchers to apply computer vision in previously inaccessible environments, while minimising the need for laborious hand-labelling and highly controlled experimental setups.

Check out Fabians recent talks from his personal website; https://fabianplum.com/

Free registration via Pretix is now open!

For more information regarding the ARC TechSocial Series, including future dates and speakers, please visit our website

Interested in other activities, events and support offered by ARC? Then head to ARC's Community & Events 

Map