license: mit
language:
- en
base_model:
- google/vit-base-patch16-224
Research paper: Segmentation of Wheat Rust Disease Using Co-Salient Feature Extraction
can be accessed via https://doi.org/10.3390/agriengineering7020023
RustCoSeg is a is a novel two-stage pipeline that first classifies wheat leaf images using a Vision Transformer, followed by segmentation using a Co-Salient Object Detection-inspired architecture, effectively isolating rust-infected regions across related samples.
Base models include: --ViT based classification and DCFM based co-salient feature extraction for segmenting out wheat rust disease.
ViT
Vision Transformer.
DCFM
The official repo of the paper Democracy Does Matter: Comprehensive Feature Mining for Co-Salient Object Detection.
Environment Requirement
create enviroment and intall as following:
pip install -r requirements.txt
Data Format
trainset: CoCo-SEG
testset: CoCA, CoSOD3k, Cosal2015
Put the CoCo-SEG, CoCA, CoSOD3k and Cosal2015 datasets to DCFM/data as the following structure:
DCFM
├── other codes
├── ...
│
└── data
├── CoCo-SEG (CoCo-SEG's image files)
├── CoCA (CoCA's image files)
├── CoSOD3k (CoSOD3k's image files)
└── Cosal2015 (Cosal2015's image files)
Trained model
trained model can be downloaded from papermodel.
Run test.py for inference.
The evaluation tool please follow: https://github.com/zzhanghub/eval-co-sod
Usage
Download pretrainde backbone model VGG.
Run train.py for training.
Prediction results
The co-saliency maps of DCFM can be found at preds.
Reproduction
reproductions by myself on 2080Ti can be found at reproduction1 and reproduction2.
reprodution by myself on TITAN X can be found at reproduction3.
Others
The code is based on GCoNet. I've added a validation part to help select the model for closer results. This validation part is based on GCoNet_plus. You can try different evaluation metrics to select the model.