XIANG-Shuai/GWFSS-competition

Introduction

Competition Page

If you want any update on the Global Wheat Dataset Community, go on https://www.global-wheat.com/

Wheat is a cornerstone of global food security, serving as a dietary staple for billions of people worldwide. Detailed analysis of wheat plants can help scientists and farmers cultivate healthier, more resilient, and more productive crops. The Global Wheat Full Semantic Segmentation (GWFSS) task aims to perform pixel-level segmentation of plant components, including leaves, stems, and heads, to comprehensively describe plant architecture, health, and development.

Semantic segmentation requires laborious and costly pixel-level annotations, which and hinder model scalability and accessibility. In this competition, we focus on the following key question: "How can we leverage a large number of unlabeled images along with a small amount of labeled data to train an effective segmentation model?"

Dataset Composition

• Splits: Pretraining, Train, Validation
• Pretraining data: Over 64,000 images from 9 different domains, resolution: 512×512 pixels
• Training data: Supervised fine-tuning data—99 images from 9 domains (11 images per domain), resolution: 512×512 pixels
• Validation data: Used for model evaluation—99 images from 9 domains (11 images per domain), resolution: 512×512 pixels. Submit predictions to CondaBench to obtain mIoU scores.

Usage

You can directly train your segmentation model using the Training Data. However, to improve the performance, we recommend first leveraging the Pretraining Data for unsupervised learning, and then using the the Training Data for supervised fine-tuning.

After training your segmentation model, you can submit your results on Codabench. Submission details can be found in the Codabench competition guidelines.

Data Structure

Pretraining Data

gwfss_competition_pretrain/
├── domain1/   # Unlabeled images from domain 1  
├── domain2/   # Unlabeled images from domain 2  
...
├── domain9/   # Unlabeled images from domain 9  

Training Data

gwfss_competition_train/
├── images/     # Original images  
├── masks/      # Colored semantic segmentation masks for visualization  
├── class_id/   # Grayscale images where pixel values represent category IDs (usable for training)  

Validation Data

gwfss_competition_val/
├── images/     # Original images  

Labels

• Format: PNG images, where pixel values correspond to category IDs
• Classes:

[
    {
        "class_id": 0,
        "name": "background",
        "color": [0, 0, 0]
    },
    {
        "class_id": 1,
        "name": "head",
        "color": [132, 255, 50]
    },
    {
        "class_id": 2,
        "name": "stem",
        "color": [255, 132, 50]
    },
    {
        "class_id": 3,
        "name": "leaf",
        "color": [50, 255, 214]
    }
]