ODDGen - Object Detection Dataset Generator

Generate object detection datasets from individual objects with automatic bounding box annotations

Features

• Automatically generates object detection dataset along with bounding box annotations
• Background remover for objects with backgrounds along with training script for U2NET if one requires a custom background remover
• Add backgrounds to generated images
• Easily scalable to generate tens of thousands of images through batching

Installation

git clone https://github.com/pranftw/oddgen.git
cd oddgen
# create and activate virtual environment
pip install -r requirements.txt

Example

• A simple ShapeNet example is demonstrated in example.py
• Download sample dataset from https://drive.google.com/drive/folders/1mGMq5jGSY_BmnOdYjejDg_yi8zeSmOc2?usp=sharing and sample backgrounds from https://drive.google.com/drive/folders/135ozo_lTppYSaquCMWP566fEpl0MyWff?usp=sharing
• The sample dataset contains 3 objects - phone, ship and plane from ShapeNet. Sample backgrounds contains several background images
• Set the required path environment variables as specified in SET ENV VARS
• Execute python example.py by uncommenting only the required functions one by one

Download PyTorch files for Background Remover

• mkdir oddgen/bg_remover/models
• Download the required files from the following links into oddgen/bg_remover/models
  • model scripts - https://drive.google.com/drive/folders/1CckJQwVC2gRu9xw0HNNw5JYhC4RHfMHQ?usp=sharing
  • ISNET weights - https://drive.google.com/file/d/1nV57qKuy--d5u1yvkng9aXW1KS4sOpOi/view?usp=sharing (from official link)
  • U2NET weights - https://drive.google.com/file/d/1ao1ovG1Qtx4b7EoskHXmi2E9rp5CHLcZ/view?usp=sharing (from official link)
  • U2NETP weights - https://drive.google.com/file/d/1rbSTGKAE-MTxBYHd-51l2hMOQPT_7EPy/view?usp=sharing (from official link)