Dataset Augmentation Workflow for Tensorflow Object Detection API

Introduction

This repository provides a comprehensive workflow for augmenting and randomizing both images and their corresponding labels (annotations generated using labelImg). The aim is to prepare a well-structured dataset for training with the Tensorflow Object Detection API.

Directory Structure

Create the following organized directory structure:

.
└───annotations
|   └───xml
└───data
│   └───split
└───images
|   └───train
|   └───test
└───sample_images
└───sample_labels
│   AnnotationAugmentation.py
│   Augmentation.py
│   Labeling.py
│   Randomize.py
│   Readme.md
│   Renaming.py
│   TrainTestSplit.py
│   XMLtoCSV.py

Steps

1. Prepare Sample Images and Labels:

   • Place raw images in the sample_images folder.
   • Utilize labelImg to label the sample images and save annotations in the sample_labels folder.

2. Image Augmentation:

   • Apply image augmentations using the script Augmentation.py.

3. Label Augmentation:

   • Apply label augmentations using the script AnnotationAugmentation.py.

4. Consolidation:

   • Copy the contents of the sample_images folder to the images folder.
   • Copy the contents of the sample_labels folder to the annotations/xml folder.

5. Randomization:

   • Shuffle the images and labels within the images and annotations/xml folders using Randomize.py.

6. Update Annotations:

   • Update XML files' contents with updated image and label directories by running Labeling.py.

7. XML to CSV Conversion:

   • Convert all XML files to a single CSV using XMLtoCSV.py.

8. Train-Test Split:

   • Create train and test subfolders within the images directory.
   • Split the dataset into train and test images using TrainTestSplit.py.

Results

• The images directory is organized into train and test subdirectories, each containing their respective images for training and testing.
• The data directory holds a split subdirectory with two CSV files for labels corresponding to train and test images.
• All images and labels are sequentially named and their directories are updated as part of the workflow.

By following this structured workflow, you can efficiently augment, randomize, and prepare your dataset for optimal utilization with the Tensorflow Object Detection API.