SmartHarvest: Multi-Species Fruit Ripeness Detection Dataset

Dataset Description

SmartHarvest is a comprehensive multi-species fruit ripeness detection and segmentation dataset designed for precision agriculture applications. The dataset contains high-resolution images of fruits in natural garden environments with detailed polygon-based instance segmentation annotations and ripeness classifications.

Key Features

• 8 fruit species: Apple, cherry, cucumber, strawberry, tomato, plum, raspberry, pepper
• Multi-class ripeness: Ripe, unripe, spoiled, plus obscured category
• Instance segmentation: Polygon annotations with 3-126 vertices per instance
• Real-world conditions: Natural lighting, occlusion, and clustering challenges
• Expert validation: Agricultural specialist annotation review and quality control

Dataset Statistics

• Total images: 486 high-resolution images
• Total annotations: 6,984 individual fruit instances
• Average annotations per image: 14.4 instances
• Polygon complexity: 14.1 ± 9.8 vertices per annotation
• Occlusion coverage: 53.8% partially obscured instances
• Image resolution: Resized and padded to 1200×1200 pixels

Supported Tasks

Primary Tasks

• Object Detection: Fruit localization with species and ripeness classification
• Instance Segmentation: Pixel-level fruit boundary delineation
• Multi-class Classification: Combined species and ripeness state prediction

Agricultural Applications

• Robotic Harvesting: Automated fruit picking with quality assessment
• Yield Prediction: Crop monitoring and harvest optimization
• Quality Control: Post-harvest sorting and grading
• Precision Agriculture: Species-specific crop management

Dataset Structure

Data Fields

Each sample contains:

{
    'image': PIL.Image,                    # Original fruit image
    'image_id': int,                       # Unique image identifier
    'annotations': [
        {
            'id': int,                     # Unique annotation ID
            'category_id': int,            # Species-ripeness category
            'species': str,                # Fruit species name
            'ripeness': str,               # Ripeness state
            'bbox': [x, y, width, height], # Bounding box coordinates
            'segmentation': [[x1,y1, ...]], # Polygon vertices
            'area': float,                 # Annotation area in pixels²
            'iscrowd': bool,               # Multiple objects flag
            'visibility': str              # Occlusion status
        }
    ],
    'metadata': {
        'source': str,                     # Image source information
        'capture_conditions': str,         # Lighting and environment
        'quality_score': float            # Annotation quality metric
    }
}

Category Mapping

Category ID	Species	Ripeness	Description
0	background	-	Background class
1	apple	unripe	Green/immature apples
2	apple	ripe	Harvest-ready apples
3	apple	spoiled	Overripe/damaged apples
4	cherry	unripe	Immature cherries
5	cherry	ripe	Harvest-ready cherries
6	cherry	spoiled	Overripe cherries
7	cucumber	unripe	Small/immature cucumbers
8	cucumber	ripe	Harvest-ready cucumbers
9	cucumber	spoiled	Overripe cucumbers
10	strawberry	unripe	White/green strawberries
11	strawberry	ripe	Red strawberries
12	strawberry	spoiled	Overripe strawberries
13	tomato	unripe	Green tomatoes
14	tomato	ripe	Red tomatoes
15	tomato	spoiled	Overripe tomatoes

Additional species (plums, raspberries, peppers) in development

Dataset Splits

Current Distribution

• Total: 486 images with 6,984 annotations
• Apple subset: 98 images, 2,582 annotations
• Cherry subset: 86 images, 969 annotations
• Tomato subset: 94 images, 1,572 annotations
• Strawberry subset: 111 images, 1,397 annotations
• Cucumber subset: 97 images, 464 annotations

Recommended Splits

For reproducible experiments, we recommend:

• Training: 80% (389 images)
• Validation: 20% (97 images)
• Stratification: Balanced across species and ripeness states

Data Collection and Annotation

Collection Methodology

• Sources: Natural garden environments, orchard partnerships
• Geographic coverage: Multiple growing regions to reduce bias
• Temporal coverage: Different seasons and growth stages
• Lighting conditions: Natural outdoor lighting with time-of-day variation
• Image quality: High-resolution captures with professional equipment

Annotation Protocol

• Tool: VGG Image Annotator (VIA) with custom configuration
• Annotators: Trained computer vision researchers with agricultural consultation
• Quality control: 25% overlap for inter-annotator agreement (κ > 0.85)
• Expert review: 10% agricultural specialist validation
• Polygon precision: Minimum 8 vertices, detailed boundary delineation

Species-Specific Criteria

Color-Based Ripeness (Apples, Tomatoes, Cherries, Peppers)

• Ripe: >75% characteristic color coverage
• Unripe: <25% color development
• Spoiled: Brown/black discoloration, visible mold

Size-Based Ripeness (Cucumbers, Pears)

• Ripe: 80-100% of variety-specific size range
• Unripe: <80% expected size
• Spoiled: Yellowing, soft spots, wrinkled skin

Texture-Based Ripeness (Strawberries, Raspberries)

• Ripe: Uniform color, firm but yielding texture
• Unripe: White/green areas, hard texture
• Spoiled: Soft spots, mold, collapsed structure

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load complete dataset
dataset = load_dataset("TheCoffeeAddict/SmartHarvest")

# Load specific split
train_data = load_dataset("TheCoffeeAddict/SmartHarvest", split="train")

# Access sample
sample = dataset['train'][0]
image = sample['image']
annotations = sample['annotations']

PyTorch Integration

import torch
from torch.utils.data import Dataset
from torchvision import transforms
from datasets import load_dataset

class SmartHarvestDataset(Dataset):
    def __init__(self, split="train", transform=None):
        self.dataset = load_dataset("TheCoffeeAddict/SmartHarvest", split=split)
        self.transform = transform
        
    def __len__(self):
        return len(self.dataset)
        
    def __getitem__(self, idx):
        sample = self.dataset[idx]
        image = sample['image']
        target = {
            'boxes': torch.tensor(sample['bboxes']),
            'labels': torch.tensor(sample['labels']),
            'masks': torch.tensor(sample['masks'])
        }
        
        if self.transform:
            image = self.transform(image)
            
        return image, target

# Usage
transform = transforms.Compose([
    transforms.Resize((800, 800)),
    transforms.ToTensor(),
])

dataset = SmartHarvestDataset(split="train", transform=transform)

Data Visualization

import matplotlib.pyplot as plt
import numpy as np

def visualize_sample(sample):
    image = sample['image']
    annotations = sample['annotations']
    
    fig, ax = plt.subplots(1, 1, figsize=(12, 8))
    ax.imshow(image)
    
    for ann in annotations:
        # Draw bounding box
        x, y, w, h = ann['bbox']
        rect = plt.Rectangle((x, y), w, h, fill=False, color='red', linewidth=2)
        ax.add_patch(rect)
        
        # Add label
        species = ann['species']
        ripeness = ann['ripeness']
        ax.text(x, y-5, f"{species}-{ripeness}", color='red', fontsize=10)
    
    ax.set_title("SmartHarvest Sample Annotation")
    plt.show()

# Visualize first sample
sample = dataset['train'][0]
visualize_sample(sample)

Baseline Results

Model Performance (Apple-Cherry Subset)

Trained Mask R-CNN with ResNet-50 backbone:

Metric	Value	Description
[email protected]	22.49%	Average precision at IoU=0.5
[email protected]	7.98%	Average precision at IoU=0.75
COCO mAP	60.63%	Mean AP across IoU 0.5-0.95

Per-Class Performance

Class	[email protected]	Notes
Apple-Ripe	10.45%	Challenging due to color variation
Apple-Unripe	25.00%	Better defined characteristics
Apple-Spoiled	32.60%	Distinctive visual features
Cherry-Ripe	18.20%	Small size challenges
Cherry-Unripe	17.10%	Consistent with apple pattern
Cherry-Spoiled	31.56%	Best performance per species

Code available at: https://github.com/Maksim3l/SmartHarvest

Considerations for Use

Strengths

• Real-world applicability: Natural garden conditions with authentic challenges
• Multi-species coverage: Broad agricultural applicability
• Expert validation: Agricultural specialist involvement in annotation
• Detailed annotations: Polygon-level segmentation for precise localization
• Ripeness granularity: Practical quality assessment categories

Limitations

• Geographic bias: Limited to specific growing regions
• Seasonal bias: Collection timing affects ripeness distribution
• Equipment bias: Single camera system characteristics
• Scale limitations: Limited images per species for production deployment
• Class imbalance: Varying representation across ripeness states

Recommended Applications

• Research benchmarking: Computer vision method evaluation
• Algorithm development: Detection and segmentation model training
• Educational use: Agricultural computer vision teaching
• Prototype development: Proof-of-concept agricultural systems

Usage Considerations

• Data augmentation: Recommended for training robustness
• Cross-validation: Stratified splits to maintain species balance
• Evaluation metrics: Use agricultural-relevant metrics beyond standard CV measures
• Deployment testing: Validate on target agricultural environments

Ethical Considerations

Data Privacy

• Image sources: Publicly available images or consent-obtained private collections
• Location privacy: No GPS coordinates or specific farm identifiers included
• Farmer consent: Proper permissions obtained for orchard data collection

Bias and Fairness

• Geographic diversity: Active efforts to include multiple growing regions
• Seasonal representation: Multiple collection periods to reduce temporal bias
• Equipment standardization: Documentation of capture conditions for bias awareness

Environmental Impact

• Sustainable agriculture: Supporting precision farming for reduced resource use
• Technology access: Open-source approach for global accessibility
• Local adaptation: Encouragement of regional dataset development

Citation

If you use this dataset in your research, please cite:

@inproceedings{loknar2025comprehensive,
    title={Comprehensive Multi-Species Fruit Ripeness Dataset Construction: From Eight-Species Collection to Focused Apple-Cherry Detection},
    author={Loknar, Maksim and Mlakar, Uroš},
    booktitle={Student Computing Research Symposium},
    year={2025},
    organization={University of Maribor},
    url={https://huggingface.co/datasets/TheCoffeeAddict/SmartHarvest}
}

Dataset Card Contact

Authors: Maksim Loknar, Uroš Mlakar
Institution: Faculty of Electrical Engineering and Computer Science, University of Maribor, Slovenia
Email: [email protected], [email protected]
Project Page: https://github.com/Maksim3l/SmartHarvest

For questions about dataset usage, additional species requests, or collaboration opportunities, please open an issue in the GitHub repository or contact the authors directly.