First Place Solution to the MLCAS 2025 GWFSS Challenge: The Devil is in the Detail and Minority

This repository contains our team(HUST_TinySmart)'s first place solution of the Global Wheat Full Semantic Segmentation.

Our solution is based on Guided Distillation. We also integrates the SAPA feature upsampling operator and utilizes ViT-Adapter as backbone networks within a semi-supervised training framework.

For details, see the paper: First Place Solution to the MLCAS 2025 GWFSS Challenge: The Devil is in the Detail and Minority

Songliang Cao, Tianqi Hu, Hao Lu

Correspondence to: hlu@hust.edu.cn, songliangcao@126.com

National Key Laboratory of Multispectral Information Intelligent Processing Technology School of Artificial Intelligence and Automation Huazhong University of Science and Technology, China.

Overview

Our solution includes three stages: in stage one, we leverage the labeled training dataset to train a supervised baseline ViT-Adapter and enhance its detail delineation with a dynamic upsampler SAPA; in stage two, we apply a semi-supervised learning pipeline with guided distillation on both labeled data and selected unlabeled data; in stage three, we implement a form of test-time scaling by zooming in images and segmenting twice following the sliding-window-style inference.

Installation

conda create -n gwfss python=3.8 -y
conda activate gwfss
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
 
# build MSDformableAttention
cd ops
sh make.sh
 
# build detectron2
cd ../projects/detectron2
pip install -e .
 
# build SAPA operators
cd ../sapa/sapa
python setup.py develop

Usage

To test our model on the GWFSS validation set, follow these instructions:

• Download our trained model from this link.
• Modify the inference.py to change the model path and data path.
• Run sh test.sh.

Here is the results of our solution on GWFSS competition:

Backbone	#Param.	Public Leaderboard	Private Leaderboard
BEiTv2-L	348.7M	0.77	0.75

Training

1. Prepare the pretrained weight

• you can download the BEiTv2 pretrained model from here

• convert the weights to d2 format:

  python tools/convert-pretrained-model-to-d2.py weight.pth weight_stage1.pkl stage1 
      
  python tools/convert-pretrained-model-to-d2.py weight.pth weight_stage2.pkl stage2

2. Parepare the dataset

• labeled data: modify both data/data/datasets/bultin.py and projects/detectron2/detectron2/data/datasets/builtin.py (refer to this issue)

• unlabled data: modify the data/datasets/gwfss_images.py; For GWFSS unlabeled data, you can select part of samples via unlabeled_4500.txt

3. Run the training：

• stage1: Supervised Training

  bash stage1_train.sh

• stage2: Guided Distillation

  bash stage2_train.sh

Citation

If you find this work or code useful for your research, please consider giving a star and citation:

@article{songliang2025gwfss
   title={First Place Solution to the MLCAS 2025 GWFSS Challenge: The Devil is in the Detail and Minority},
   author={Cao, Songliang and Hu, Tianqi and Lu, Hao},
   booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
   year={2025}
}

License

Codes and model weights are released under MiT license. See LICENSE for additional details.