DCT

This repository includes the official implementation of the paper:

Dynamic Color Transform Networks for Wheat Head Detection

Plant Phenomics, 2022

& Dynamic Color Transform for Wheat Head Detection

International Conference on Computer Vision Workshops (ICCVW), 2021

Chengxin Liu, Kewei Wang, Hao Lu, Zhiguo Cao

Huazhong University of Science and Technology, China

Highlights

• DCT is intuitive and steerable
• We won 2nd Place in Global Wheat Challenge 2021 based on DCT
• DCT is also applicable to other vision tasks, e.g., crowd counting

Installation

• Set up environment

# env
conda create -n dct python=3.7
conda activate dct

# install pytorch
conda install pytorch==1.10.0 torchvision==0.11.0 -c pytorch -c conda-forge

• Install

# clone 
git clone https://github.com/cxliu0/DCT.git
cd DCT

# install dependecies
pip install -r requirements/build.txt

Data Preparation

• Download Global Wheat Head Detection Dataset 2021 (GWHD 2021). We expect the directory structure to be as follows:

DCT
├── data
│    ├── gwhd_2021
│         ├── images
│         ├── competition_test.csv
│         ├── competition_train.csv
│         ├── competition_val.csv
├── configs
├── models
├── ...

• Preprocess data

python preprocess_data.py

• Download pretrained yolov4-p7 model and put it into ./pretrained/

Training

We follow a two-step training strategy.

• Step 1: train a baseline Scaled-YOLOv4 model

sh train_baseline.sh

• Step 2: fix the weights of the baseline model, and train DCT:

sh train_dct.sh

Evaluation

We implement the evaluation of ADA following Global Wheat Challenge 2021 (GWC 2021). Note that the tested ADA in this repository seems slightly lower than the official one.

• Modify test_ADA.sh

weights: modify it to the path of your locally trained model

• Run

sh test_ADA.sh

Pretrained models (Tested on the GWHD 2021 Dataset)

We also provide pretrained DCT models. You can download the models if you do not want to train DCT.

DCT Type	DCT Arch	Val ADA	Test ADA	Weights
Regression	ResNet18	0.787	0.629	regDCT.pt
Classification	ResNet18	0.782	0.630	clsDCT.pt

Citation

If you find this work or code useful for your research, please consider citing:

@article{liu2022dct,
  title = {Dynamic Color Transform Networks for Wheat Head Detection},
  author = {Chengxin Liu  and Kewei Wang  and Hao Lu  and Zhiguo Cao },
  journal = {Plant Phenomics},
  year={2022}
}

@INPROCEEDINGS{liu2021dct,
  author={Liu, Chengxin and Wang, Kewei and Lu, Hao and Cao, Zhiguo},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)}, 
  title={Dynamic Color Transform for Wheat Head Detection}, 
  year={2021},
  pages={1278-1283},
}

Acknowledgment

This repository is based on Scaled-YOLOv4.