Related Experiment Video
Updated: Aug 7, 2026
11:37
RGB and Spectral Root Imaging for Plant Phenotyping and Physiological Research: Experimental Setup and Imaging Protocols
Published on: August 8, 2017
Low-cost monocular RGB-based 3D structural mapping for horticultural plants via semantic scene completion
Ruijun Jing1,2, Rui An3, Zhuoxing Li3
1School of Instrument and Electronics, North University of China, Taiyuan, China.
Frontiers in Plant Science
|August 6, 2026
Summary
This study introduces a monocular 3D mapping framework for precision agriculture, enabling detailed crop structural analysis from single RGB images. The system provides accurate canopy height and volume, supporting autonomous operations and crop monitoring.
Area of Science:
- Computer Vision
- Robotics
- Agricultural Science
Background:
- Precision agriculture requires detailed 3D structural data (canopy height, volume) for crop monitoring and operational safety.
- Existing LiDAR or RGB-D sensor methods are costly and impractical for large-scale horticultural environments.
- Current 2D and SLAM-based pipelines yield insufficient semantic maps for agricultural structural analysis.
Purpose of the Study:
- To develop a monocular 3D structural mapping framework for horticultural plants using semantic scene completion.
- To enable actionable structural analysis from single RGB images, overcoming limitations of existing methods.
- To derive task-oriented maps like canopy height, volume, and traversability layers.
Main Methods:
- Proposes a monocular 3D mapping framework via semantic scene completion for horticultural plants.
- Introduces a Depth-Aware Decoder Module for explicit depth recovery and 2D-to-3D feature fusion.
- Employs an NCS-Guided Geometry Encoder and Global Encoder for depth-aware global relational modeling and local detail refinement.
- Utilizes an occupancy head for 3D semantic completion and generates structural maps from RGB input.
Main Results:
- Achieved 82.31% occupancy IoU, 84.26% mIoU, and 86.25% precision on a custom horticultural dataset.
- Demonstrated low mean absolute error (MAE) for canopy height (0.019-0.026 m) and relative errors for canopy volume (8.4%-11.4%) via field measurements.
- Validated effectiveness on both Semantic-KITTI and the custom horticultural dataset, with the deployed model being RGB-only.
Conclusions:
- The proposed RGB-only monocular 3D structural mapping framework effectively provides actionable insights for crop monitoring.
- The method overcomes limitations of costly sensors and sparse semantic maps, enabling robust structural analysis in agriculture.
- Results show significant potential for enhancing autonomous robotic operations and crop management in horticultural settings.
