Core Concepts
Neural Radiance Fields offer accurate phenotyping in greenhouses, comparable to traditional methods but with improved scalability and robustness.
Abstract
This study explores the use of Neural Radiance Fields for accurate 3D phenotyping of pepper plants in greenhouses. Traditional phenotyping methods are compared to NeRF reconstruction, showing competitive accuracy. The study addresses challenges in current phenotyping methods and introduces a method for recovering the true scale of NeRF. A detailed methodology is provided for data acquisition, 3D modeling from NeRF, point cloud registration, and post-processing for phenotyping measurements. Results show that NeRF models achieve similar accuracy to 3D scanning methods with improved efficiency and accuracy after scale restoration. The study also includes an ablation study on NeRF-based approaches, demonstrating improvements in model optimization and reconstruction quality. Quantitative evaluations on accuracy and phenotypic measurements show promising results.
Structure:
Introduction
Background on Plant Phenotyping
Traditional vs Implicit Modeling Approaches
Methodologies: Data Acquisition, 3D Modeling from NeRF, Point Cloud Registration, Post-processing
Experiment and Discussion: Experimental Setup, Ablation Study on NeRF-based Approach, Quantitative Evaluation on Accuracy, Demonstration on Phenotypic Measurements
Conclusion and Future Work
Stats
The mean distance error between the scanner-based method and the NeRF-based method is 0.865mm.
Quotes
"Neural Radiance Fields offer accurate phenotyping in greenhouses."
"The study shows that learning-based NeRF method achieves similar accuracy to 3D scanning-based methods."
"Neural Reconstruction by Neus improves surface reconstruction quality."