Photorealistic Reconstruction of the Sterkfontein Caves Using Novel View Synthesis
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of the Witwatersrand, Johannesburg
Abstract
Neural Radiance Fields (NeRF) have emerged as a groundbreaking approach for novel view synthesis (NVS), inspiring numerous variants aimed at enhancing robustness and fidelity. A notable advancement in this domain is RAW image training, which addresses the challenges of reconstructing low-light, noisy images. However, most existing studies have been evaluated on datasets captured in controlled environments and rely on abundant computational resources — conditions that severely limit the practicality of NeRF-based methods in real-world applications. In response, this work introduces a novel, resource-efficient pipeline tailored for challenging low-light scenarios, exemplified by our newly captured datasets from the Sterkfontein Caves. These datasets are not only significant for palaeontological and speleological research but also serve as a rigorous testbed for evaluating NVS methods under extreme conditions. Our pipeline builds on state-of-the-art techniques while incorporating strategic modifications — such as refined hyperparameter tuning and alternative output activation functions — to mitigate issues like gradient explosion and geometry collapse. Experimental results demonstrate that our approach achieves high-fidelity reconstructions even in suboptimal lighting conditions, thereby paving the way for more practical and accessible neural rendering solutions in resource-constrained environments.
Description
A dissertation submitted in partial fulfilment of the requirements for the degree of Master of Science, to the Faculty of Science, School of Computer Science & Applied Mathematics, University of the Witwatersrand, Johannesburg, 2025
Keywords
Citation
Liu, Ireton. (2025). Photorealistic Reconstruction of the Sterkfontein Caves Using Novel View Synthesis. [Master's dissertation, University of the Witwatersrand, Johannesburg]. WIReDSpace. https://hdl.handle.net/10539/50316