hic-yolov5's Introduction

HIC-Yolov5

This repo contains the code for the HIC-YOLOv5, the original paper is:

HIC-YOLOv5: Improved YOLOv5 For Small Object Detection
Shiyi Tang, Fang Yini, Shu Zhang
[Paper]

To get started

1. Requirements

Run pip install -r requirements.txt in terminal.

2. Prepare Visdrone-2019 dataset

(a) You can download the dataset from https://github.com/VisDrone/VisDrone-Dataset#task-1-object-detection-in-images.

(b) Convert data form to Yolo by running visDrone2yolov5.py (you may need to change the dir).

We suppose the data directory is constructed as

Your project name
├── datasets
|   ├── VisDrone2019
|   |   └── VisDrone2019-DET-train
            └── annotations
            └── images
            └── labels
|   |   └── VisDrone2019-DET-val
            └── annotations
            └── images
            └── labels
|   |   └── VisDrone2019-DET-test-dev
            └── annotations
            └── images
            └── labels
├── yolov5-6.0

3. Train the model

Modify args in train.py. In HIC-Yolov5, some args are set as follows:

--weights: yolov5s.pt
-cfg: models/yolov5s-p2-involution-cbam.yaml
--hyp: data/hyps/hyp.scratch-high.yaml

4. Evaluate the model

Modify args in val.py.

--weights: the best.pt file in your result folder.
--task: can be val or test.

Cite

If you find this work useful in your research, please cite the paper:

@misc{tang2023hicyolov5,
      title={HIC-YOLOv5: Improved YOLOv5 For Small Object Detection}, 
      author={Shiyi Tang and Yini Fang and Shu Zhang},
      year={2023},
      eprint={2309.16393},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}