---

Synthetic Data Generation Tool

This repository contains tools for generating synthetic images by combining objects with various backgrounds to create diverse datasets for machine learning training.

• ⚠️ Make sure to adjust the image scaling to match the background image sizes.
• Adjust the code block here : scale_factor = random.uniform(0.35, 0.55)
• Make sure that all the product images are in PNG format for easy detection of the alpha leveles (https://pixlr.com/remove-background/)

scale_factor = random.uniform(0.35, 0.55)

Project Overview

This tool is designed to solve a critical challenge in computer vision development: the need for large, diverse datasets of object images in various real-world contexts. Here's what makes this project powerful:

Key Features

• Automated Image Generation: Combines object images with various backgrounds to create realistic, synthetic training data
• Multi-Processing Support: Offers both single and multi-process modes for efficient image generation
• Smart Alpha Channel Handling: Uses transparency information for precise object placement
• Flexible Dataset Creation: Supports multiple object categories (no limits) and various background scenarios
• Automatic Labeling: Generates labels suitable for machine learning training pipelines

Use Cases

• Training object detection models
• Developing computer vision systems
• Creating synthetic datasets
• Testing object placement algorithms
• Generating synthetic data for environment simulation

Benefits

• Reduces the need for manual photo shoots
• Creates consistent, well-labeled datasets
• Enables rapid testing with different environmental conditions
• Saves time and resources in dataset creation
• Ensures diverse training data for better AI model performance

Sample Workflow

For Object Detection

Input

Background Image

Product Images

Output

For Image Segmentation

Setup

1. Clone the repository:

git clone https://github.com/HGSChandeepa/Synthetic-data-generation.git

2. Create and activate a virtual environment (optional but recommended):

python -m venv .venv

# Activate (each new terminal)
# macOS/Linux:
source .venv/bin/activate
# Windows CMD:
.venv\Scripts\activate.bat
# Windows PowerShell:
.venv\Scripts\Activate.ps1

3. Install required dependencies:

pip install -r requirements.txt

4. Prepare your data structure:

   • Place background images in the backgrounds/ folder
   • Place product images in their respective numbered folders under products/
   • Ensure your dataset configuration is set in dataset/data.yaml

Directory Structure

├── backgrounds/          # Background images
├── products/             # Product images (organized by categories 1-20)
├── dataset/              # Dataset configuration and labels
├── images/
├── detection/
├── segmentation/              # Sample images
└── visualization/        # Visualization outputs

Image Labeling Mechanism

The system uses alpha channels for precise product placement and masking:

1. Alpha Channel Usage:

   • Product images should have transparent backgrounds (alpha channel)

   • The alpha channel is used to:

     • Determine exact product boundaries
     • Create precise masks for placement
     • Maintain product integrity during compositing

2. Product Placement:

   • Products are placed on backgrounds using their alpha channels
   • This ensures clean edges and natural-looking integration
   • No manual masking required due to alpha channel automation

Running the Program

1. Single Process Mode:

python main.py

2. Multi-Process Mode (for faster processing):

python multiple_processes_main.py

3. Validation:

python validate.py

Output

• Generated images will be saved in the specified output directory
• Each image combines a background with one or more products
• Labels are automatically generated for AI training

Notes

• Ensure product images have transparent backgrounds
• Background images should be high resolution
• Check data.yaml for dataset configuration