High-fidelity wheat plant reconstruction using 3D Gaussian splatting and neural radiance fields

Area of Science:

• Computer Vision
• Plant Science
• Robotics

Background:

• Traditional 3D plant reconstruction methods are often sparse or noisy.
• View synthesis models offer high accuracy but require extensive data and precise calibration.
• Accurate 3D plant models are crucial for advanced phenotyping.

Purpose of the Study:

• To develop a comprehensive dataset for 3D plant reconstruction using view synthesis.
• To evaluate the performance of 3D Gaussian splatting (3DGS) and neural radiance fields (NeRF) for plant modeling.
• To provide accessible tools and data for 3D plant phenotyping research.

Main Methods:

• A novel robotic camera capture system was developed.
• A multiview dataset of 20 wheat plants across 15 weeks was created.
• Two view synthesis models, 3DGS and NeRF, were trained and evaluated.

Main Results:

• Both 3DGS and NeRF generated high-fidelity plant reconstructions from novel viewpoints.
• Accurate 3D plant point cloud representations were achieved with millimeter-level accuracy.
• The system demonstrated effective 3D modeling of plants over their growth cycle.

Conclusions:

• View synthesis methods like 3DGS and NeRF are transformative for 3D plant phenotyping and reconstruction.
• The released dataset and software will advance research in active vision and plant modeling.
• This work facilitates more accurate and detailed analysis of plant structures.