Kunitomi/coffee-bean-maskrcnn

Coffee Bean Detection - Fine-tuned Mask R-CNN

Model Details

Model Description

This is a fine-tuned Mask R-CNN model specialized for detecting and segmenting individual coffee beans in images. The model performs instance segmentation, providing precise pixel-level masks and bounding boxes for each detected coffee bean.

• Developed by: Mark Kunitomi
• Model type: Instance Segmentation (Mask R-CNN)
• Architecture: ResNet-50 FPN backbone
• License: apache-2.0

Example Results

Left: Original green coffee beans image. Right: Detection and segmentation results showing individual bean masks with confidence scores.

How to Get Started with the Model

Quick Start with Jupyter Notebook

A complete inference notebook (inference_demo.ipynb) is included with this model.

Command Line Interface

For quick inference, use the included predict_beans.py script:

# Install dependencies
pip install -r requirements.txt

# Basic inference on a single image
python predict_beans.py --model maskrcnn_coffeebeans_v1.safetensors --images your_image.jpg

# Process multiple images with custom settings
python predict_beans.py --model maskrcnn_coffeebeans_v1.safetensors \
  --images *.jpg \
  --confidence 0.5 \
  --nms_threshold 0.3 \
  --smooth_polygons \
  --filter_edge_beans \
  --output_dir results \
  --export_format coco

# For all available options
python predict_beans.py --help

Key CLI Options:

• --confidence: Detection confidence threshold (default: 0.5)
• --nms_threshold: Non-maximum suppression threshold (default: 0.3)
• --smooth_polygons: Apply polygon smoothing for cleaner masks
• --filter_edge_beans: Remove beans touching image edges
• --export_format: Output format (json, coco, labelme, all)

Recommendations

• Use confidence threshold of 0.5 or higher for production applications
• Validate on your specific bean varieties before deployment
• Consider additional fine-tuning for specialized use cases
• Implement human verification for critical applications

Training Details

Training Data

The model was trained on a custom dataset of coffee bean images:

• 128 training images with detailed COCO-format annotations
• Multiple coffee varieties and roast levels
• Various lighting conditions and backgrounds
• Manually annotated polygon masks for each bean
• Data augmentation: rotation, scaling, color jittering, horizontal/vertical flips

Preprocessing

• Images resized to maintain aspect ratio
• Normalization with ImageNet statistics
• Random augmentations during training

Evaluation

Testing Data

Evaluated on a held-out validation set:

• 4,952 ground truth bean instances
• Diverse bean arrangements and densities
• Various roast levels and lighting conditions

Metrics

Metric	Value
Precision	99.92%
Recall	96.71%
Average IoU	90.93%
Detection Rate	96.71%
Average Confidence	99.82%
Mask Loss	0.1333
Validation Loss	0.2464

Technical Specifications

Model Architecture and Objective

• Architecture: Mask R-CNN with ResNet-50 Feature Pyramid Network
• Input: RGB images (any size)
• Output: Instance masks, bounding boxes, class labels, confidence scores
• Objective: Minimize combined classification, box regression, and mask segmentation losses
• Model Size: 176.1 MB (SafeTensors format)

Compute Infrastructure

Hardware

• CPU: Mac Mini M2 with 8GB RAM
• Training time: ~2 hours for fine-tuning

Software

• PyTorch 2.0+
• TorchVision 0.15+
• CUDA 11.8+
• Python 3.8+