yudo's Introduction

YOLO for Uniform Directed Object detection

This project is an implementation of the paper YUDO: YOLO for Uniform Directed Object Detection. The codebase is an adaptation of the popular YOLOv7 model, used for detection of directed objects with uniform dimensions.

Requirements

docker build --rm --no-cache -t yudo:version_1 -f Dockerfile .

docker run --gpus device=0 --rm --shm-size=1G -ti -v {YOUR CODE PATH}:/yudo --name yudo yudo:version_1

Install in addition:

pip install 'git+https://github.com/facebookresearch/detectron2.git'

Prepare annotations

The dataset used in this project is obtained from Honeybee Segmentation and Tracking Datasets. An image cropping and adapting labels' format to the yolo format can be done using the gen_yolo_anns.py script.

Run the training

A training command example:

python train.py \
--epochs 200 \
--workers 4 \
--device 0 \
--batch-size 2 \
--data data/data.yaml \
--img-size 512 512 \
--cfg cfg/training/yolov7-tiny.yaml \
--weights 'yolov7-tiny.pt' \
--name model_001 \
--hyp data/hyp.scratch.yaml \
--image-weights \
--exist-ok \
--adam

Citation

If you find this project usefull, please consider citing the paper:

@article{nedeljkovic2023yudo,
  title={YUDO: YOLO for Uniform Directed Object Detection},
  author={Nedeljkovi{\'c}, {\DJ}or{\dj}e},
  journal={arXiv preprint arXiv:2308.04542},
  year={2023}
}

yudo's People

Contributors

Stargazers

Watchers

yudo's Issues

License

Hi djordjened92,

congratulations on your excellent result, it's a real breakthrough in the field of oriented objects detection, which until now has been dominated by algorithms that find only rotated bounding box and not the direction. Would you mind adding any open source license to your code so that more researchers can legally use your work in their experiments?

Recommend Projects