Semantic Segmentation Competition

This repository contains code and results for a semantic segmentation competition using a small autonomous driving dataset. The project explores various data augmentation, network architectures, and training strategies to improve segmentation accuracy and efficiency.

Dataset

• Training images: 150
• Testing images: 50
• Classes: 19 (see seg_data/names.txt)
• Color palette: see seg_data/colors.txt

Features

• PyTorch-based data loading and augmentation
• Custom semantic segmentation network with options for BatchNorm, Residual, Dropout, etc.
• Multiple data augmentation strategies (flip, color jitter, crop, rotation, blur, etc.)
• Training and evaluation scripts with mIoU metric
• FLOPs calculation for model efficiency
• Ablation study and results tracking

Usage

1. Setup

   • Install dependencies (PyTorch, torchvision, numpy, PIL, etc.)
   • (Optional) Use Google Colab or Kaggle for GPU acceleration

2. Prepare Data

   • Place the dataset in the seg_data/ directory as described above.

3. Run Experiments

   • Open and run segmentation-2025.ipynb or any notebook in the result/ subfolders for specific experiments.
   • Modify data augmentation, network architecture, and training parameters as needed.

4. Evaluate

   • Evaluation is performed using mean Intersection over Union (mIoU).
   • Results and predictions are saved in the result/ subfolders.

5. Report

   • See Final_Report.pdf for a summary of methods, results, and ablation studies.

Example: Running the Baseline

Open segmentation-2025.ipynb and run all cells to train and evaluate the baseline model.

Results

• Baseline mIoU: ~0.28
• Best mIoU (with augmentation and advanced architecture): see ablation tables in the report and notebooks

References

• PyTorch Documentation
• fvcore FLOPs Analysis

---

Authors:
Ge Wang, Tien

For more details, see the code and Final_Report.pdf.