InspectGaussian: Large-scale coarse-to-fine Gaussian reconstruction for orchard inspection robots

Area of Science:

• Robotics and Computer Vision
• Agricultural Technology
• 3D Reconstruction

Background:

• Large-scale 3D reconstruction for orchard environments is crucial for robotic inspection and precision agriculture.
• Existing methods face challenges with unstructured scenes, variable lighting, and computational demands.

Purpose of the Study:

• To develop an efficient, coarse-to-fine Gaussian reconstruction framework (InspectGaussian) for orchard inspection robots.
• To improve the accuracy and detail of 3D reconstructions in complex agricultural settings.

Main Methods:

• Utilized an RGB-D data acquisition strategy with ORB-SLAM3 and a dense mapping module for pose estimation and point cloud generation.
• Employed a divide-and-conquer approach: YOLO-World for plant view extraction and 3D matching, followed by improved 3D Gaussian Splatting (3DGS) for plant-specific reconstruction.
• Incorporated depth regularization and region-aware refinement into the 3DGS method.

Main Results:

• Achieved 96% precision and 93% recall in plant view extraction.
• Outperformed state-of-the-art methods in reconstruction fidelity (31.226 PSNR, 0.915 SSIM, 0.067 LPIPS) and point cloud accuracy (7 mm error).
• Demonstrated effectiveness in capturing fine structural and textural details with scalability and efficiency.

Conclusions:

• InspectGaussian provides a practical solution for high-throughput, in-field plant phenotyping.
• The framework lays the foundation for intelligent orchard monitoring and management.
• The method is effective for detailed 3D reconstruction in challenging agricultural environments.