TS_OAMIL-for-Plant-disease-detection

• This code is an implementation of our manuscript: An Iterative Noisy Annotation Correction Model for Robust Plant Disease Detection.

• The manuscript is currently under peer review. We submitted it to Frontiers in Plant Science.

• Authors: Jiuqing Dong, Alvaro Fuentes, Sook Yoon*, Hyongsuk Kim*, Dong Sun Park

• We will release the code and complete this documentation as soon as possible.

Installation

• Set up environment

  conda create -n TSOAMIL python=3.7

  conda activate TSOAMIL

  conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge

• install dependecies

  pip install -r requirements/build.txt

• install mmcv (will take a while to process)

  cd mmcv MMCV_WITH_OPS=1 pip install -e .

  cd .. pip install -e .

• Generate noisy annotations: Please refer to files in utils/

  There are two version noise you can choose, random noise and real world noise(default gamma is 0.15).

Dataset prepare

You can download the GWHD2021 dataset from this link: https://zenodo.org/record/5092309

The original file annotation file format is '.csv'. You can transfer to a 'xml' file by using this command python ./process_gwhd/excel_to_xml.py

train

sh train.sh

test

sh test.sh

We will optimize our code and complete this documentation as soon as possible.

How to use this code for a customer dataset?

Operational steps for the teacher-student learning paradigm:

step 1: Inference the training set and save XML files. (Modify the last few lines of mmcv.visualization.image.py) (Modify the test path to the training set in configs.base.datasets.paprika_detection_oamil.py)

step 2: Perform XML to JSON conversion. Run python ./utils/xml_to_json_COCOformat_paprika.py

step 3: Generate the PKL file. Run python ./utils/gen_update_paprika.py

step 4: Modify the relevant file paths and parameters. Training path, training parameters, save path, GPU, and port.

Important notes for adding the Control class: After adding a class, you need to modify the annotation files for train, val, and test sets. Inconsistent classes between training, testing, and validation can cause errors. When generating the PKL file, modify ./utils/coco_dataset.py to adjust the classes. The number of classes in mmdet.datasets.samples.coco.py and customer.py also needs to be adjusted.