Crops3D

Util for the Crops3D Dataset

Crops3D: A Diverse 3D Crop Dataset for Realistic Perception and Segmentation

Overview

Crops3D is a 3D crop dataset designed to support research in agricultural computer vision. It is derived from real-world agricultural scenarios and aims to facilitate advanced 3D point cloud analysis tasks. The dataset is characterized by its diversity, authenticity, and complexity, providing a resource for researchers and practitioners working in the agricultural domain.

Key Features

• Diversity: Crops3D includes data from various point cloud acquisition methods and encompasses eight distinct crop types with a total of 1,230 samples.
• Authenticity: The dataset authentically represents crops in real-world agricultural settings, providing a realistic basis for research and development.
• Complexity: The intricate structures of the crops in Crops3D exhibit higher complexity compared to existing 3D public datasets, featuring substantial self-occlusion and increased complexity as crops mature.

Supported Tasks

Crops3D is designed to support three critical tasks in 3D crop phenotyping:

1. Instance Segmentation of Individual Plants: Precise segmentation of individual plants in agricultural settings.
2. Plant Type Perception: Accurate identification and classification of different crop types.
3. Plant Organ Segmentation: Detailed segmentation of plant organs, enabling fine-grained analysis.

Citation

If you use Crops3D in your research, please cite our paper:

@article{crops3d2024,
  title={Crops3D: A Diverse 3D Crop Dataset for Realistic Perception and Segmentation toward Agricultural Applications},
  author={Zhu, J. and Zhai, R. and Ren, H. et al.},
  journal={Scientific Data},
  year={2024},
  volume={11},
  number={1438},
  doi={10.1038/s41597-024-04290-0}
}

Dataset Download

The dataset is stored in the figshare database and can be accessed and downloaded using the following DOI: https://doi.org/10.6084/m9.figshare.27313272.

The dataset consists of a series of PLY and HDF5 files, compressed into a file named Crops3D.zip. It includes four main directories:

1. Crops3D Directory

• Contains annotated raw point cloud data.
• Includes eight directories named after specific crops: Cabbage, Cotton, Maize, Potato, Rapeseed, Rice, Tomato, and Wheat. Each directory stores PLY point cloud files for the respective crop.
• Two text files list the relative paths of samples in the training and test sets.

2. Crops3D_10k Directory

• Contains point cloud data subsampled to 10,000 points using FPS (Farthest Point Sampling).
• The file structure mirrors that of the Crops3D directory for consistency and ease of use.

3. Crops3D_10k-C Directory

• Contains a corrupt directory and two HDF5 (H5) files representing the training and test datasets.
• The corrupt directory includes 36 H5 files: one clean test set and 35 files corresponding to seven types of corrupted test sets, each divided into five severity levels.

4. Crops3D_IS Directory

• Contains point cloud data intended for individual plant segmentation at the plot scale (instance segmentation).
• Includes one subdirectory and two text files. The subdirectory stores point cloud data for 50 plots, covering maize, potato, and rapeseed crops, formatted in PLY.
• The two text files enumerate the relative paths of samples in the training and test sets.