This repository presents the COCO ToolKit code and tutorials for ease-of-use. COCO ToolKit provides the following options:

• Download any category from the COCO dataset
• Download any number of images per category (considering the COCO dataset limits, sure)
• Generate labels for images in the YOLO format (Mask R-CNN and One-Hot in near future)
• Training the YOLOv5 model on a custom dataset support

Installation on Windows or Linux (Tested on Windows 10, Ubuntu 18.04)

1. Clone this repository locally:

git clone https://github.com/Electronics-AI/COCO_ToolKit.git

2. Install the requirements:

pip install -r requirements.txt

3. Download and install pycocotools:

pip install -U git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

Installation in Google Colab (Tested in Google Colab)

1. Clone this repository locally:

!git clone https://github.com/Electronics-AI/COCO_ToolKit.git

2. Download and install pycocotools:

!pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

3. Install the Google Colab requirements:

!pip install -r colab_requirements.txt

1. Create your dataset folder and a categories.txt file. Let's for instance create an Animals dataset near the COCO_ToolKit folder and suppose our dataset contains 4 categories: dog, cat, bear and bird. My categories.txt file will look like this:

dog
cat
bear
bird

And here is my directory structure:

|                                               
|---Animals
|          categories.txt
|---COCO_ToolKit
               coco_toolkit/
               colab_requirements.txt
               configure.py
               LICENSE
               ...

2. Configure your dataset (You must be in the COCO_Toolkit directory)

python configure.py --dataset_path=<dataset folder path> --categories_path=<categories.txt path> --labels=<labels to generate>

For my directory structure configuration looks so:

python configure.py --dataset_path ../Animals --categories_path ../Animals/categories.txt --labels yolo

1. Let's run our code!

python run.py [--all] --train=<images per category in the train set to download> \
                      --val=<images per category in the validation set to download> \
                      --web_batch=<download images at once from COCO website>

Note: --all argument is responsible for:

• Dataset directory structure generation (labels, images, train, val folders)
• Downloading annotations from the COCO website
• Downloading images from the COCO website
• Labels generation

For my dataset I want to:

• Download 1000 images per category in the train set ("--train 1000" argument)
• Download all images per category in the validation set ("--val -1 argument", -1 means download all the images)
• I have a good internet connection, so let's set the batch of images to download = 20 ("--web_batch 20" argument)
  My run.py start looks like this:

python run.py --all --train 1000 --val -1 --web_batch 20

Wait until run.py finishes it's work and you are ready to use your custom COCO dataset!

For more information, such as training YOLOv5 in Google Colab with COCO ToolKit, see the docs folder.