Skip to main content

​
Dataset Structure

The training pipeline expects a specific directory structure as defined in data.yaml. The dataset is organized into three splits: 
train: /path/to/training/data/train/
val: /path/to/training/data/valid/
test: /path/to/training/data/test/

nc: 3
names: ['cardboard paper', 'metal', 'plastic']

​
Directory Layout

training/data/
├── train/
│   ├── images/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   │   └── ...
│   └── labels/
│       ├── image1.txt
│       ├── image2.txt
│       └── ...
├── valid/
│   ├── images/
│   └── labels/
└── test/
    ├── images/
    └── labels/
Each split (train/val/test) must contain both images/ and labels/ subdirectories. Image filenames must match their corresponding label files.

​
Class Configuration

The model is trained to detect 3 classes of recyclable materials:
Class IDClass NameDescription
0cardboard paperCardboard boxes, paper products
1metalAluminum cans, tin containers
2plasticPlastic bottles, containers

​
Image Requirements

​
Supported Formats

  • Image Types: JPG, JPEG, PNG, BMP
  • Recommended Format: JPG (for storage efficiency)
  • Color Space: RGB

​
Image Specifications

1

Resolution

Images will be resized to 640x640 during training. Higher resolution source images (1920x1080 or higher) are recommended for better quality.
2

Aspect Ratio

Any aspect ratio is supported. The training pipeline automatically handles letterboxing to maintain proportions.
3

Quality

Use clear, well-lit images. Avoid heavily compressed or blurry images that may impact model performance.

​
Best Practices

  • Diversity: Include varied backgrounds, lighting conditions, and angles
  • Occlusion: Include partially occluded objects to improve robustness
  • Scale Variation: Capture objects at different distances/sizes
  • Multiple Objects: Include scenes with multiple trash items
Avoid using images with extreme blur, heavy noise, or very low resolution (below 416x416) as they may negatively impact training.

​
Annotation Format

YOLO format uses normalized coordinates for segmentation masks. Each annotation file is a text file with the same name as its corresponding image.

​
YOLO Segmentation Format

class_id x1 y1 x2 y2 x3 y3 ... xn yn
  • class_id: Integer representing the class (0, 1, or 2)
  • x, y: Normalized polygon coordinates (0.0 to 1.0)
  • Each line represents one object instance

​
Example Annotation

training/data/train/labels/image1.txt
2 0.123 0.456 0.234 0.567 0.345 0.678 0.456 0.789 0.567 0.890
0 0.678 0.123 0.789 0.234 0.890 0.345 0.901 0.456
This example shows:
  • One plastic object (class 2) with a 5-point polygon
  • One cardboard paper object (class 0) with a 4-point polygon
Coordinates are normalized by dividing pixel values by image width (for x) and height (for y). All values must be between 0.0 and 1.0.

​
Data Augmentation

YOLOv11 includes built-in augmentation strategies that are automatically applied during training:

​
Spatial Augmentations

  • Mosaic: Combines 4 images into one for multi-scale learning
  • Random Scaling: Varies object sizes (0.5x to 1.5x)
  • Random Rotation: Up to ±10 degrees
  • Random Flipping: Horizontal and vertical flips
  • Random Translation: Shifts images within bounds

​
Pixel-level Augmentations

  • HSV Augmentation: Varies hue, saturation, and value
  • Brightness/Contrast: Random adjustments
  • Noise Addition: Adds subtle Gaussian noise

​
Advanced Techniques

  • MixUp: Blends two images and their labels
  • CutOut: Randomly masks rectangular regions
  • Perspective Transforms: Simulates different camera angles
Augmentation is only applied to the training set. Validation and test sets use original images for accurate performance evaluation.

​
Train/Val/Test Splits

SplitPercentagePurpose
Train70-80%Model learning and weight updates
Validation10-15%Hyperparameter tuning and early stopping
Test10-15%Final model evaluation

​
Split Guidelines

1

Random Distribution

Randomly distribute images across splits to ensure balanced representation.
2

Class Balance

Maintain similar class distributions in each split to prevent bias.
3

No Overlap

Ensure no image appears in multiple splits to prevent data leakage.
4

Minimum Samples

Each split should have at least 100 images per class for reliable results.

​
Dataset Validation Checklist

Before starting training, verify:
  • All images have corresponding annotation files
  • Annotation files use correct class IDs (0, 1, 2)
  • Coordinates are normalized (0.0 to 1.0)
  • Directory structure matches data.yaml paths
  • No corrupt or unreadable images
  • Balanced class distribution across splits
  • Minimum 300-500 images per class for good results
YOLOv11 will report dataset statistics at the start of training, including class distribution and any issues with annotations.

​
Next Steps

Once your dataset is prepared and validated:

Start Training

Configure and run the training script with your prepared dataset

Build docs developers (and LLMs) love

Get started for freeTalk to us