Improved You Only Look Once Image Tiling

A python script that tiles (slices) images to a desired dimensions. Here is a screenshot of sample result using tile_yolo.py.

If you're interested in knowing something for small object detection check out this gist.

Table of contents

• Usage
• Installation
• Recommended configurations
• Latest Update
• Contributing
• License

Usage

(Back to top)

1. Download tile_yolo.py.

2. Place tile_yolo.py in the desired directory where user would like tiled images to be stored.

   Note: tile_yolo.py will automatically create the directories it requires.

3. (Optional) Have a look at Recommended configurations

Installation

(Back to top)

1. Install Python (preferably, version <= 10) Download.

   Note: as of this writing, Python 3.11 or greater may cause error.

2. Download a python compatible IDE. I recommend Jupyter Notebook, PyCharm, DataSpell, or DeepNote.

   Note: I wouldn't recommend a cloud service such as Google Colab due to the large amount of images that may be generated.

3. Install the required packages listed in requirements.txt by running the following command in the terminal: pip install -r requirements.txt

Recommended configurations

(Back to top)

1. Users should have an annotated dataset. Check out Roboflow or LabelImg for quick and easy YOLO image annotation. Refer to their instructions/installation requirements.

2. When to use? Image tiling is recommended when the object of interest is very small compared to the parent image. Example: Small Object Detection: An Image Tiling Based Approach

3. tile_yolo.py will create the necessary directories to run successfully however, users are encouraged to modify the directory creation code to fit their personal needs!

4. Recommended approach: Place untiled images with their respective annotations into a folder called Input. Upon running tile_yolo.py the contents in Input will tiled to a user's desired dimensions and tiles will be placed into a folder named Images. Finally, a user can decide to move the tiled annotation (.txt) files into a separate folder named Labels or keep the tiled images and their corresponding annotations together.

   Note: Keeping the tiled images and their annotations in the same directory is recommended for YOLO - YOLOv4

   Note: Separating tiled images into Images and their annotations into Labels is recommended for YOLOv5 - YOLOv7

The python script can be run in terminal or in an IDE (PyCharm works great!).

Latest Update

(Back to top)

Added several improvements to nagi1995 original

     -Added directory-level tiling (all images in folder instead of a single image) :)

     -Added user choice to keep tiled files in one directory (Recommended for YOLOv1-YOLOv4)
            ../datasets/images/im0.jpg  # image
            ../datasets/images/im0.txt  # label

     -Added user choice to separate files into separate directories (Recommended for YOLOv5, YOLOX, YOLOR, YOLOv7, etc..)
            ../datasets/images/im0.jpg  # image
            ../datasets/labels/im0.txt  # label

     -Added the creation of Images and Labels directory (user can modify code to choose a different directory)

     -Added empty 'null' tile saving (if a tile doesn't have an annotation, user can discard or save them).

     -Added progress bars via tqdm 

Contributing

(Back to top)

Your contributions are always welcome! Special thanks to Mr. David Schmitt!