Doctoral Thesis
Image-based 3D Reconstruction of Complex Plant Structures
Murdoch University
Doctor of Philosophy (PhD), Murdoch University
2025
DOI:
https://doi.org/10.60867/00000077
Abstract
Image-based 3D reconstruction is a fundamental task in computer vision. The process involves recovering the geometry of objects using RGB images. The applications of image-based 3D reconstruction span across many domains, including virtual reality, gaming, autonomous driving and fashion. 3D reconstruction of plants has benefits in phenotyping, crop and growth monitoring, robotic pruning, and remote operations. Accurate plant geometry reconstruction using images to automate processes in agriculture is crucial and remains challenging. The thesis aims to develop image-based deep learning methods to overcome challenges in plant reconstruction.
The thesis has several major contributions. We explored the strengths and limitations of image-based depth estimation methods and discussed the challenges in the reconstruction of plant structures, such as leaves, stems, and flowers, with wind-induced motion. We identified that rapid, non-rigid motions of complex plant structures caused reconstruction methods to fail. We proposed a Neural Radiance Fields (NeRF) based framework to model the images of a plant with moving structures into a static canonical form and decomposed the canonical scene into foreground and background, thus enabling the network to focus on movements.
We tackled two known limitations of NeRF-based methods. To overcome the scene-specificity and longer training times, we proposed a novel plane-based representation that learns to represent multiple static and dynamic scenes. Explicit feature planes were used to represent the view space volume and a shared canonical volume across multiple scenes. Finally, we aggregated the previous findings and proposed a single image-based framework to reconstruct multiple static or dynamic plants or multiple growth stages of a single plant. The framework enabled edge-preserving, surface-aware learning for plant structures with signed distance functions.
The results showed the effectiveness of implicit and explicit neural representations in motionaware geometry reconstruction of plants. The thesis provides key insights, including the importance of modeling the rapid motions in canonical space, the benefits of a multi-scene representation, surface-aware learning, and the use of thin features on image-based neural representation for complex plant structures.
Details
- Title
- Image-based 3D Reconstruction of Complex Plant Structures
- Authors/Creators
- Rajapaksha Pathiranage Uchitha Nilakshi Rajapaksha
- Contributors
- Ferdous Sohel (Supervisor) - Murdoch University, School of Information TechnologyHamid Laga (Supervisor) - Murdoch University, Centre for Biosecurity and One HealthDean Diepeveen (Supervisor) - Murdoch University, School of Information Technology
- Awarding Institution
- Murdoch University; Doctor of Philosophy (PhD)
- Publisher
- Murdoch University
- Identifiers
- 991005862489807891
- Murdoch Affiliation
- School of Information Technology
- Resource Type
- Doctoral Thesis
Metrics
7 File views/ downloads
16 Record Views