Datasets:

SamruddhK
/

coffee-bean-grading-dataset

Automated quality assessment of raw agricultural products is critical for ensuring fair trade and supply chain efficiency. This dataset presents 3,877 high-resolution images of pre-roast Arabica coffee beans, collected from farms in **Coorg, Karnataka (India)**—a major coffee-producing region. Each bean is categorized into one of four quality grades:

Grade	Description
A	Premium
B	Good
C	Standard
D	Defective

A total of 2,284 beans are annotated using polygon masks and grade labels in LabelMe JSON format, supporting research in classification, instance segmentation, automated grading, and defect detection. A YOLOv11 multi-class instance segmentation baseline was trained to validate annotation quality and demonstrate practical model performance.

Dataset Structure

/
├── CGA/   # Grade A (Premium)
│   ├── CGA_images/
│   └── CGA_json/
├── CGB/   # Grade B (Good)
├── CGC/   # Grade C (Standard)
└── CGD/   # Grade D (Defective)

Distribution Summary

Grade	Quality	Images	Polygon Annotations
A	Premium	1000	600
B	Good	1000	600
C	Standard	1002	487
D	Defective	875	597
Total	—	3,877	2,284

Data Collection and Annotation

Attribute	Details
Coffee Variety	Coffea robusta
Region	Coorg (Kodagu), Karnataka, India
Imaging Setup	Controlled indoor lighting with a white backdrop
Devices Used	iPhone 15 Pro, Samsung 2023/24 models, Google Pixel 7, Poco X Series
Annotation Tool	LabelMe (Polygon Mode)
Data Format	`.jpg` images + `.json` polygon annotation files

Each annotation file contains:

"label": "grade_a" | "grade_b" | "grade_c" | "grade_d"
"points": [ [x1,y1], [x2,y2], ... ]

Annotations were reviewed by trained annotators to ensure precision.

Recommended Use Cases

Multi-class instance segmentation training
Automated grading and sorting systems
Agricultural defect detection research
Food quality assurance studies
Robust low-cost supply chain inspection systems

Out-of-Scope Use

Personal identification
Medical or biometric inference (dataset contains no personal data)

Baseline Experiment (YOLOv11 Segmentation)

To validate the dataset's quality and visual separability, a baseline instance segmentation model (YOLOv11) was trained. The model was trained for 150 epochs, with the best-performing checkpoint saved for evaluation.

The results confirm that the dataset supports strong performance for automated grading, achieving a mean Average Precision (mAP) of 0.93 for object detection and 0.91 for instance segmentation.

Final Metrics (from `best.pt` model)

Metric	Grade A	Grade B	Grade C	Grade D	Mean (All Classes)
Box mAP@0.5	0.94	0.90	0.91	0.97	0.93
Mask mAP@0.5	0.92	0.88	0.88	0.96	0.91

Note: The high precision on Grade D (Defective) is particularly valuable, as it demonstrates the model's reliability in identifying and sorting out low-quality beans, which is a primary goal of automated grading systems.

How to Use

Load JSON Annotation Example

import json
import glob

# Example for Grade A
files = glob.glob("CGA/CGA_json/*.json")
with open(files[0], "r") as f:
    ann = json.load(f)

print(ann["shapes"][0]["points"])  # Polygon coordinates
print(ann["shapes"][0]["label"])   # Grade label

Value of the Dataset

First publicly available polygon-annotated dataset of pre-roast coffee beans.
Enables end-to-end automated grading using segmentation + classification.
Facilitates fair pricing and quality transparency in the coffee supply chain.
Robust for deployment in low-cost rural environments using consumer smartphones.

Contributors

Name	ORCID	Role
Samruddh K	0009-0008-3588-9272	Research, Dataset Preparation & Documentation
Abhay Varun S	0009-0003-1299-724X	Research, Dataset Collection & Annotation
Bopanna K N	0009-0008-0432-3196	Annotation Support & Verification
H A Dheemanth Gowda	0009-0001-1891-632X	Annotation Support & Verification

Citation

If you use this dataset, please cite:

@dataset{pre_roast_coffee_grading_2025, title = {A Image Dataset of Pre-Roast Arabica Coffee Beans with Polygon Annotations for Automated Grading}, author = {Samruddh K and Bopanna K N and H A Dheemanth Gowda and Abhay Varun S}, year = {2025}, publisher = {Hugging Face Datasets}, license = {MIT}, url = {https://huggingface.co/datasets/SamruddhK/coffee-bean-grading-dataset} }