🚜 Offroad Semantic Segmentation

An advanced semantic segmentation pipeline optimized for autonomous off-road navigation in unstructured environments. Developed for the Cosmos-2026 Hackathon.

---

👥 Team: CodeVizards

• Team Name: CodeVizards
• Team Members: T Mohamed Yaser, Mohammed Sabeehuddin

---

🚀 Overview

Navigating off-road terrain is a complex challenge for autonomous systems. Unlike urban environments with clear lanes and signs, off-road paths are unstructured, yielding a high degree of variability in texture and geometry.

Our solution leverages a DeepLabV3+ architecture with a ResNet-50 backbone, specifically tuned for the FalconCloud digital twin environment.

✨ Key Features

• Hybrid Objective Function: Combines Categorical Cross-Entropy (0.7) and Dice Loss (0.3) to maximize Mean IoU (mIoU).
• Sim2Real Robustness: Heavy spatial and color augmentations using Albumentations to bridge the gap between simulation and real-world textures.
• High-Resolution Training: Optimized 384px tiling strategy for preserving fine-grained geometric boundaries.
• Performance Optimization:
  • AMP (Automated Mixed Precision) for 2x faster training.
  • Test-Time Augmentation (TTA) for stable inference.

---

📊 Results & Visualization

🛠 Model Performance

Our model demonstrates high stability across diverse terrains (rocks, mud, trails, vegetation).

Metric	Value
Model	DeepLabV3+ (ResNet-50)
mIoU Target	> 70%
Framework	PyTorch

🖼 Inference Sample Target

Above: A sample prediction showing accurate segmentation of off-road terrain classes.

📈 Training Progress Target

Convergence shown across epochs using our hybrid loss strategy.

---

🛠 Tech Stack

• Core: PyTorch, TorchVision
• Augmentation: Albumentations
• Data Handling: CV2, NumPy, tqdm
• Environment: Google Colab / TPU v5e

---

⚙️ Usage

1. Clone the repository:

   git clone https://github.com/Yaser-123/offroad-semantic-segmentation.git
   cd offroad-semantic-segmentation

2. Install Dependencies:

   pip install torch torchvision albumentations opencv-python tqdm

3. Run the Notebook: Open cosmos.ipynb in your favorite Jupyter environment (Colab/VS Code) and follow the step-by-step pipeline from data loading to final inference.

---

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

_{Developed with ❤️ by CodeVizards for Cosmos-2026}