Pantazis, Omiros;
(2024)
Data-Efficient Computer Vision for Biodiversity Monitoring.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Pantazis_10200299_Thesis.pdf Download (62MB) | Preview |
Abstract
Accurate and efficient biodiversity monitoring at a global scale is critical for informed decision-making and meeting sustainable development goals. Although recent deep learning solutions excel on computer vision benchmarks, their benefits are not directly transferable to biodiversity data, which present challenges like reliance on expert annotations, noise, environmental variability, class imbalance, and fine-grained categorization. This thesis explores how computer vision techniques can be efficiently adapted and advanced to enhance biodiversity monitoring despite these challenges. First, we demonstrate that deep learning classifications effectively align with expert annotations to inform ecological findings (Chapter 2). We quantify their robustness across various data and model manipulations, highlighting the importance of prioritizing training set size and quality. This motivates Chapter 3, where we explore self-supervised learning to reduce the need for extensive human-provided labels. We show that self-supervision is effective for species classification from camera trap images and propose a novel method leveraging spatiotemporal context that consistently outperforms baselines in low-data scenarios. Subsequently, we examine large vision-language models for tackling ecological tasks with few or even no labels (Chapter 4). We demonstrate these models' shortcomings and propose a self-supervised method to adapt them for real-world tasks, including biodiversity monitoring. The results show that our method effectively adapts vision-language models with limited labels, even for challenging biodiversity monitoring tasks. In the final chapter (Chapter 5), we address natural world image retrieval, requiring both advanced image understanding and domain expertise. We introduce INQUIRE, a text-to-image retrieval benchmark featuring a large set of wildlife images and expert-level queries. Benchmarking on INQUIRE highlights limitations of current vision-language models in handling fine-grained queries, emphasizing the need for improved methods to maximize the value extracted from wildlife images. Overall, this thesis advances data-efficient computer vision, addressing key challenges in biodiversity monitoring and enhancing the capabilities of automated systems for ecological applications.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Data-Efficient Computer Vision for Biodiversity Monitoring |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10200299 |
Loading...
Loading...Loading...Loading...Archive Staff Only
 |
View Item |
