Tomato-Nerf: Advancing Tomato Model Reconstruction With Improved Neural Radiance Fields

Abstract

The real-time simulation of large-scale agricultural operations will offer farmers data-driven and physically consistent decision support, facilitated by predictive digital twins. To construct a predictive digital twin, the initial step involves 3D reconstruction of plant geometry. In this paper, a high-resolution, accurate 3D reconstruction of tomato plants, Tomato-NeRF, is proposed, which is specially used for three-dimensional reconstruction of tomato plants. Our approach used a modular design to integrate ideas from their research paper into Tomato-NeRF. By using hash encoding to map coordinates to trainable feature vectors, we balance quality, memory usage, and performance in NeRF training. The proposal sampler targets key regions for rendering, and customized loss functions are designed to optimize specific tasks. The effectiveness of our approach is demonstrated by the ability to generate high-resolution geometric models from phone camera data. Comparative results show that Tomato-NeRF has significant advantages over Instant-NGP and MipNeRF in the tomato plant reconstruction task. The data acquisition method is simpler and more efficient than other reconstruction methods, providing a practical solution for real-time agricultural simulations.

Keywords

• 3D reconstruction
• agricultural automation
• deep learning
• NeRF
• multi-view imaging