Aaltoes-CV-Hackthon: Surf in Wild Rift

This repository contains a PyTorch-based pipeline for AI-image Detection tasks on a Kaggle competition dataset. It demonstrates how to train and evaluate popular segmentation models.

---

🏆 This project achieved a 97.2% score and secured 2nd place in the Aaltoes 2025 Computer Vision v1 Hackthon.
👥 Team: Ziyi Wang, Rijie Hao, Zhuchenyang Liu
🙏 Special thanks to the organizers at Aaltoes for hosting this amazing challenge.

📚 Table of Contents

1. ✨ Features
2. 💻 Requirements
3. 🗂️ Dataset
   • Overview
   • Files
   • Data Format
4. 🚀 Usage
   • Training
   • Inference
   • Threshold Optimization
5. 🤝 Model Ensemble
6. 📊 Results
7. ⚙️ Arguments and Options
8. 📁 Project Structure
9. 🔗 Related Resources
10. 📝 License

---

Features

• Multiple models in one framework: Easily switch between U-Net, ResUNet, AttentionUNet, SegFormer, and Mask2Former.
• Flexible device setup: Runs on GPU if available; falls back to CPU automatically otherwise.
• Configurable hyperparameters: Adjust learning rate, batch size, epochs, threshold, etc.
• Run-Length Encoding (RLE): Automatically encodes predicted masks in RLE format for Kaggle submissions.
• Validation: Splits the training dataset into train/val sets for monitoring performance via Dice coefficient.
• Easy inference: Generate submission files for Kaggle evaluation.

---

Requirements

• PyTorch (with GPU support if you want to train on CUDA)
• torchvision
• OpenCV
• scikit-learn
• pandas
• tqdm

You can install the required Python packages via:

pip install -r requirements.txt

---

Dataset

You can download dataset via

kaggle competitions download -c aaltoes-2025-computer-vision-v-1

Please place your data in Datasets folder.

Overview

This competition provides a dataset of 46836 images, each with its original version and corresponding binary mask indicating inpainted regions. All images are 256×256 pixels in size.

Files

• train/ - Directory containing training data:
  • images/ - Manipulated versions of the images
  • masks/ - Binary masks indicating inpainted regions
  • originals/ - Original versions of the images
• test/ - Directory containing test data:
  • images/ - Manipulated versions of the images
• sample_submission.csv - Example submission file in the correct format

Data Format

Images

• Format: PNG
• Resolution: 256×256 pixels
• Color space: RGB (3 channels)

Masks

• Format: PNG
• Resolution: 256×256 pixels
• Values:
  • 0: Original (non-inpainted) regions
  • 1: Inpainted (manipulated) regions

---

Usage

Training

To train a model, run the main script with appropriate arguments. You can also you the run.sh to train or inference.

python train.py \
    --model "Mask2Fomer" \
    --batch_size 16 \
    --epochs 30 \
    --lr 1e-4 \
    --num_workers 4 \
    --train_image_dir "../Dataset/train/train/images" \
    --train_mask_dir "../Dataset/train/train/masks" \
    --test_image_dir "../Dataset/test/test/images" \
    --threshold 0.5

Inference

For inference you need add model name and the path to the pretrained model.

python train.py \
    --model "Mask2Former" \
    --test_image_dir "../Dataset/test/test/images" \
    --load_pretrain "/path/to/your/best_model.pth" \
    --threshold 0.5

---

Threshold Optimization

Validation Set Inference

First, generate the float mask predictions from your target model on the validation set:

python train.py \
    --model "Mask2Former" \
    --predict_validation True \
    --load_pretrain "/path/to/your/best_model.pth" 

Optimization

Next, you can optimize the threshold using Mask2Former/threshold_optimization.ipynb. To adjust the search gap, modify the following line:

thresholds = np.linspace(0.4, 0.7, 20) 

Example output:

Optimal threshold: 0.637 with Dice coefficient: 0.9667

Threshold results:
    Threshold  Mean_Dice
0    0.400000   0.392585
1    0.415789   0.392585
2    0.431579   0.392585
3    0.447368   0.392585
...

---

Models Ensemble

To further improve the prediction score, try the ensemble method.

After gathering multiple encoded submission.csv by different models, you can try the ensemble strategy using Mask2Former/ensemble.ipynb.

Adjust the combination of models and the voting threshold by modifying: VOTE_THRESHOLD.

---

🏆 Our Result

The final private leaderboard result is based on the ensemble of three fine-tuned models with threshold optimization:

Model	Threshold	Public Dice Coefficient	Private Dice Coefficient
mask2former-swin-small-ade-semantic	0.61	0.9678	0.9651
mask2former-swin-base-ade-semantic	0.62	0.9686	0.9677
mask2former-swin-large-ade-semantic	0.61	0.9659	0.9639
Ensemble	/	0.9733	0.9713

Here are some prediction samples:

🎉 We achieved second place in the Aaltoes-CV-Hackathon competition!

License

Shield:

This project is licensed under a MIT License.

---

Related resources

We acknowledge all the open-source contributors for the following projects to make this work possible:

• Mask2Former

• SegFormer

• IML-ViT

• Huggingface 🤗