Skip to main content

​
Overview

The Reciclaje AI model has been trained on a diverse dataset of waste imagery, focusing on common recyclable materials found in household and commercial waste streams. The training data ensures the model can accurately identify and classify different types of waste materials.
The model was trained specifically for the 5 core waste categories: Metal, Glass, Plastic, Carton, and Medical waste.

​
Dataset Composition

​
Waste Categories

The training dataset includes images across all 5 detection classes:
Examples include:
  • Aluminum cans (soda, beer)
  • Metal food containers (tin cans)
  • Metal bottle caps
  • Small metallic objects and packaging
Images captured various sizes, conditions, and orientations of metal waste items.
Examples include:
  • Glass bottles (clear, green, brown)
  • Glass jars and containers
  • Broken glass pieces
  • Various glass packaging materials
Training data includes different glass colors and transparency levels.
Examples include:
  • PET bottles (water, soft drinks)
  • Plastic containers and packaging
  • Plastic bags and film
  • Various plastic product types
Includes different plastic types, colors, and forms commonly found in waste.
Examples include:
  • Cardboard boxes (various sizes)
  • Paper cartons (milk, juice)
  • Corrugated cardboard
  • Paperboard packaging
Training images include both clean and slightly damaged cardboard materials.
Examples include:
  • Disposable masks
  • Medical gloves
  • Healthcare-related waste items
  • Sanitization materials
Focused on common medical waste items found in general waste streams.

​
Data Characteristics

​
Image Variety

The training dataset was designed to include:
  • Multiple Angles: Objects photographed from different perspectives
  • Varying Lighting: Images captured under different lighting conditions
  • Different Backgrounds: Various environmental contexts
  • Mixed Conditions: Clean and dirty/damaged items
  • Scale Variation: Objects at different distances and sizes

​
Real-World Scenarios

The model was trained on real-world waste imagery to ensure practical applicability in actual recycling and waste management scenarios.
Training data reflects common scenarios such as:
  • Items in recycling bins
  • Waste on conveyor belts
  • Mixed waste situations
  • Individual items against various backgrounds

​
Data Preprocessing

Standard YOLOv8 preprocessing techniques were applied: 
# Typical preprocessing steps
- Image resizing and normalization
- Data augmentation (rotation, scaling, brightness adjustment)
- Bounding box annotation
- Train/validation split

​
Dataset Limitations

While comprehensive, the training data has some limitations:
Rare Items: The model may not recognize waste items that were rarely or never included in the training set.

​
Underrepresented Scenarios

Certain conditions may be underrepresented:
  • Extreme lighting conditions (very dark or bright)
  • Unusual perspective angles
  • Heavily damaged or degraded items
  • Uncommon waste item variations
  • Items with significant occlusion
For detailed information about model limitations, see the Limitations page.

​
Continuous Improvement

The model can be improved through:
  • Additional Training Data: Expanding the dataset with more diverse examples
  • Fine-tuning: Adjusting the model for specific use cases or environments
  • Active Learning: Incorporating feedback from real-world deployments

​
Contributing Data

If you have waste imagery that could improve the model:
1

Collect Images

Gather clear, well-lit images of waste items from the 5 categories.
2

Annotate

Label bounding boxes and assign correct class IDs.
3

Share

Consider contributing to the open-source project or training your own fine-tuned version.

​
Technical Details

​
Annotation Format

The model uses YOLO format annotations: 
<class_id> <x_center> <y_center> <width> <height>
Where coordinates are normalized (0-1) relative to image dimensions.

​
Class Mapping

clsName = ['Metal', 'Glass', 'Plastic', 'Carton', 'Medical']

# Class ID to name mapping:
# 0 -> Metal
# 1 -> Glass
# 2 -> Plastic
# 3 -> Carton
# 4 -> Medical

​
Next Steps

Model Overview

Learn about the model architecture and specifications

Limitations

Understand known limitations and edge cases

Build docs developers (and LLMs) love

Get started for freeTalk to us