Jurassic model: active learning with self-supervised transformers for efficient 3d fossil segmentation
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of the Witwatersrand, Johannesburg
Abstract
Fossil research is vital for understanding Earth’s biodiversity and evolutionary history, yet traditional excavation and manual segmentation methods are time-consuming and prone to human error. Our research addresses these challenges by presenting a novel semi-automated pipeline that employs advanced computer vision techniques, specifi- cally a 3D segmentation model enhanced with a transformer backbone. By retraining the model on out-of-domain medical data and employing active learning strategies, our approach significantly reduces the reliance on extensive labelled datasets, achieving a segmentation accuracy that is markedly superior to conventional 2D models. Empirical results demonstrate that the proposed method streamlines the segmentation process and enhances the identification of critical fossil features with minimal human intervention. Our findings indicate that the integration of uncertainty sampling allows for a similar accuracy level to fully labelled datasets while utilising less than 20% of the required labels. This research contributes to the field of computer vision by providing a robust framework for non-destructive analysis of fossil scans, with broader implications for other domains that struggle with limited access to high-quality labelled data. The outcomes of this work suggest that our framework could revolutionise fossil studies and promote further advancements in automated segmentation technologies across various scientific disciplines.
Description
A research report submitted in fulfillment of the requirements for the Master of Science, in the Faculty of Science, School of Computer Science and Applied Mathematics, University of the Witwatersrand, Johannesburg, 2025
Citation
Boyley, Andrew. (2025). Jurassic model: active learning with self-supervised transformers for efficient 3d fossil segmentation [Master’s dissertation, University of the Witwatersrand, Johannesburg]. WIReDSpace. https://hdl.handle.net/10539/50281