YOLO_Underwater

This is the official code for our paper "ULO: An Underwater Light-weight Object Detector for Edge Computing", If you are interested in our work, please consider citing the following:

@Article{machines10080629,
AUTHOR = {Wang, Lin and Ye, Xiufen and Wang, Shunli and Li, Peng},
TITLE = {ULO: An Underwater Light-Weight Object Detector for Edge Computing},
JOURNAL = {Machines},
VOLUME = {10},
YEAR = {2022},
NUMBER = {8},
ARTICLE-NUMBER = {629},
URL = {https://www.mdpi.com/2075-1702/10/8/629},
ISSN = {2075-1702},
DOI = {10.3390/machines10080629}
}

Introduction

This repo is based on

• YOLO-v3
• pjreddie-darknet
• YOLO-v5
• YOLO Nano
• YOLOX
• GhostNet

Project Structure

│  main.py
│  README.md
│  test.py
│  train.py
│  val.py
│
├─data
|      image
|      box
│      test.txt
│      train.txt
│      urpc.names
│      val.txt
│
├─dataloader
│      data_split.py
│      URPCDataset.py
│
├─models
│  │  basic_layers.py
│  │  darknet_model.py
│  │  ghost_module.py
│  │  preprocessing_module.py
│  │  select_model.py
│  │  yolo_nano.py
│  │  yolo_nano_underwater.py
│  │  yolo_underwater.py
│  │  yolo_underwater_tiny.py
│  │
│  └─cfg
│          yolov3-tiny.cfg
│          yolov3.cfg
│          yolov4-tiny.cfg
│          yolov4.cfg
│
└─utils
        compute_anchor.py
        logger.py
        opts.py
        stats.py
        utils.py

Installation

git clone git@github.com:wangsssky/YOLO_Underwater.git
pip install -r requirements.txt

Dataset

A optimized version of URPC2019 is used in the work, the updated annotations are available at https://github.com/wangsssky/Refined-training-set-of-URPC2019.

Train & Evaluate

train

python main.py --model YOLO-Underwater-Tiny --image_size 512  --num_epochs 300 
--batch_size 64 --lr 0.001 --num_threads 64 --gpu --weight_decay 5e-4 --preprocessing  
--checkpoint_path ./ckpt_YOLO_Underwater 

test

python main.py --model YOLO-Underwater --image_size 512  --batch_size 1 
--num_threads 4 --gpu  --test True --no_train --no_val --preprocessing
--resume_path ckpt_YOLO_Underwater/best.pth --conf_thresh 0.25 --nms_thresh 0.45