Object Detection with COCO_OI and ObjectNet_D

This repository provides an implementation for object detection using two complementary datasets to COCO: COCO_OI and ObjectNet_D. These datasets can help to improve object detection performance by overcoming the limitations of COCO dataset saturation.

Getting Started

To get started with the environment and dependencies, follow these steps:

1. Create a Conda Environment

Create a new Conda environment with Python 3.7:

conda create -n objDet python=3.7

2. Activate the Conda Environment

Activate the newly created environment:

conda activate objDet

3. Install Dependencies

Install the necessary Python packages using pip:

pip install pycocotools numpy==1.16.0 opencv-python tqdm tensorboard tensorboardX pyyaml webcolors matplotlib torch==1.4.0 torchvision==0.5.0 boto3

Citation

If you use this repository or datasets in your research, please cite the following paper:

@article{borji2022complementary,
  title={Complementary datasets to COCO for object detection},
  author={Borji, Ali},
  journal={arXiv preprint arXiv:2206.11473},
  year={2022}
}

Datasets

1. COCO_OI

COCO_OI is a large dataset composed of images from COCO and OpenImages. It includes images from their 80 common classes, and it features:

• 1,418,978 training bounding boxes over 380,111 images.
• 41,893 validation bounding boxes over 18,299 images.

2. ObjectNet_D

ObjectNet_D consists of images depicting objects in daily life situations. It is originally created for object recognition tasks known as ObjectNet, and contains:

• 29 common categories with COCO.

This dataset is useful for testing the generalization ability of object detection models, especially when objects are placed in challenging and varied real-life settings.

Evaluations

This repository includes an evaluation of object detection models on these datasets. Through this evaluation, we aim to pinpoint the sources of errors and help improve the accuracy and robustness of object detection systems.

Usage

You can use this repository to train and test object detection models on COCO_OI and ObjectNet_D datasets. For more details on how to use these datasets, please refer to the code and data available at the following link:

https://github.com/aliborji/COCO_OI

---

Feel free to explore and contribute to the repository! If you have any questions or issues, don't hesitate to open an issue or reach out.