TL;DR
This paper introduces the "Leaf on Stem Image Dataset Beans" (LSID-Beans), a comprehensive dataset of 6981 images of common bean leaves. It provides detailed annotations including leaf dimensions, segmentation masks, and camera calibration parameters, enabling the development of advanced deep learning models for non-destructive plant analysis.
What Problem Does This Paper Address?
Understanding plant responses to environmental factors like soil fertility, light availability, and water stress is crucial for agriculture and plant physiology. Leaf dimensions (area, perimeter, length, width) are key indicators of these responses. However, accurately measuring these dimensions, especially non-destructively and at scale, is challenging. Existing methods often require manual labor or specialized equipment. This paper addresses the need for a robust, well-annotated dataset that can facilitate the development of automated, deep learning-based solutions for leaf dimension estimation, thereby advancing plant science research.
Methods
The authors collected images of 612 common bean leaves using a smartphone camera. Each image includes a fiducial marker (ArUco library) for scale and pose estimation. The dataset was meticulously annotated with:
- Leaf Dimensions: Actual measurements of area, perimeter, length, and width were obtained through a combination of physical measurements (ruler, curvimeter, precision balance for area based on paper grammage) and image analysis.
- Segmentation Masks: Pixels are classified as background, leaf, or marker. This was achieved through a multi-step process involving Photoshop's "Quick Selection Tool," Suzuki and Abe's contour extraction algorithm, and manual refinement using the "Basic Image Labeling Tool" and a custom tool.
- Area Maps: Each pixel in the leaf region is assigned a constant value representing the leaf's total area divided by the number of leaf pixels. Marker pixels have their area estimated via 3D reconstruction from ArUco pose.
- Camera Calibration: Intrinsic and extrinsic camera parameters were estimated using Zhang's algorithm, crucial for 3D reconstruction and pose estimation.
- Marker Pose: The 3D position and orientation of the fiducial marker were determined using the ArUco library and algorithms like Collins and Bartoli.
The data underwent significant processing, including image cropping to a 1:1 aspect ratio and resizing to 512x512 pixels, while ensuring all annotations were adjusted accordingly. A rigorous 5-fold cross-validation protocol was established to ensure balanced data distribution and prevent bias.
Datasets
The paper introduces the Leaf on Stem Image Dataset Beans (LSID-Beans).
- Content: 6981 images of 612 common bean leaves (Phaseolus vulgaris).
- Annotations: Leaf dimensions (area, perimeter, length, width), segmentation masks (background, leaf, marker), area maps, marker pose, capture conditions, and camera calibration parameters.
- Format: Images in JPG format; annotations in XML and RAW formats.
- Availability: Accessible via Mendeley Data with the identifier 10.17632/f42hwwrpgn.2.
Key Insights and Ideas
- Bridging Computer Vision and Plant Physiology: The dataset directly addresses the need for high-quality, annotated data at the intersection of computer vision and plant physiology, enabling automated analysis of plant health and response.
- Comprehensive Annotation for Multi-task Learning: The inclusion of segmentation masks, dimension annotations, and camera calibration parameters allows for the development of models capable of performing multiple tasks simultaneously (e.g., segmentation and dimension estimation).
- Robustness through Natural Backgrounds and Variability: Images were captured in natural environments with varying conditions, promoting the development of models that generalize well to real-world scenarios.
- Fiducial Marker for Accurate 3D Information: The use of ArUco markers provides a reliable method for scale, pose estimation, and potential 3D reconstruction, adding significant value beyond 2D measurements.
- Structured Cross-Validation for Reproducibility: The detailed cross-validation protocol ensures that researchers can reliably benchmark their models against this dataset.
Key Results
The paper demonstrates the utility of the LSID-Beans dataset by evaluating a DeepLabv3+ based model for leaf area estimation. The results show promising performance:
- Leaf Area Estimation: The average Relative Error Rate (RER) for leaf area estimation across all cross-validations ranged from 6.757% to 8.099%.
- Marker Area Estimation: The model achieved significantly lower RER for marker area estimation, ranging from 2.287% to 2.714%, attributed to the marker's planar geometry.
- Dataset Suitability: The results confirm the dataset's effectiveness in training models for leaf area estimation, highlighting its potential for advancing non-destructive plant analysis.
Limitations and Future Work
- Area Matrix Simplification: The current area matrices assign a constant average area value to all leaf pixels. The authors acknowledge this simplification and plan to address it in future work by estimating each leaf's 3D geometry to approximate true leaf area more accurately.
- Marker Pose Ambiguity: For some markers, pose estimation had multiple solutions, requiring manual correction and filtering based on projected area differences.
- Future Work: The authors intend to complement the dataset by estimating leaf 3D geometry. They also suggest that the proposed methods and dataset can be adapted for leaves of other species, contributing to broader applications like disease detection and species classification.
Summary updated