rgautroncgiar/croppie_coffee_ug

Croppie training dataset © 2024 by Alliance Bioversity & CIAT, Producers Direct and M-Omulimisa is licensed under CC BY-SA 4.0

Croppie training datasets

General information

Croppie dataset for machine-vision assisted coffee cherry detection. The dataset is made of a mix of Arabica and Robusta coffee tree parts (with and without a background isolation element) with individual bounding boxes around all coffee cherries.

The original dataset is composed of 633 images with about 61 050 unique bounding boxes over coffee cherries in YOLO format. This original dataset has been processed to cut-down all images into 480 x 640 size pieces and the full original image downscaled to 480 x 640. We provide the processed dataset with Python scripts that allow easy visualization of the annotated dataset.

Coffee cherries of more than 10mm (following the longitudinal axis) are annotated according to their color:

• green
• yellow
• red
• dark brown (overripe/dry cherries)
• an extra class indicates low visibility/unsure label appreciation.

Here is an example of an annotated image:

Data structure

This repository has the following structure:

.
├── annotation_guide.html # original annotation instructions
├── classes.json  # json to convert numerical classes into the cherry type
├── data
│   ├── train.zip
│   └── val.zip
├── images
│   ├── annotated_1688033955437.jpg
│   ├── train_counts.png
│   └── val_counts.png
├── README.md
└── scripts  # script for easy visualization of the annotated data
    ├── label_training_images.py
    └── requirements.txt

Dataset information

Each numerical class corresponds to the following cherry type:

{0: "dark_brown_cherry", 1: "green_cherry", 2: "low_visibility_unsure", 3: "red_cherry", 4: "yellow_cherry"}

• train:
  • Training dataset
  • 5 836 annotated images
  • YOLO format

• val:
  • Validation dataset
  • 2 497 annotated images
  • YOLO format

• annotation_guide.html: instructions provided to label the images for cherry detection

Scripts

The script label_training_images.py allows to label the images of the datasets and saves them in a folder ./_labelled_dataset_images. Assuming you are in the scripts folder, first run pip3 install -r requirements.txt if required package are not installed. After that, simply run python3 label_training_images.py

License

Croppie training dataset © 2024 by Alliance Bioversity & CIAT, Producers Direct and M-Omulimisa is licensed under CC BY-SA 4.0