Offroad Semantic Segmentation

Duality AI Code Sprint Hackathon – Offroad Autonomy Track

Semantic segmentation for desert environment understanding

---

📋 Table of Contents

• Project Overview
• System Requirements
• Environment Setup
• Project Structure
• Training
• Testing & Inference
• Expected Outputs
• Troubleshooting
• Hackathon Compliance

---

🎯 Project Overview

This project implements a semantic segmentation model for off-road autonomous navigation in desert environments. The model segments images into 10 semantic classes relevant for path planning and obstacle avoidance.

Key Features

• DINOv2 Backbone: Leverages self-supervised vision transformer features
• Multi-Scale Decoder: Progressive upsampling for precise segmentation
• Mixed Precision Training: FP16 for memory efficiency
• Robust Pipeline: Automatic handling of missing data and edge cases

Model Performance

Metric	Value
Best mIoU	27.69%
Best Dice	35.62%
Accuracy	67.23%

---

💻 System Requirements

Minimum Requirements

Component	Specification
OS	Windows 10/11, Linux, macOS
Python	3.8 or higher
RAM	8 GB
GPU	NVIDIA GPU with 4GB+ VRAM
CUDA	11.8 or higher
Storage	5 GB free space

Recommended Requirements

Component	Specification
GPU	NVIDIA RTX 3050 or better (8GB+ VRAM)
RAM	16 GB
Storage	SSD with 10 GB free space

---

🛠 Environment Setup

Option A: Automated Setup (Recommended)

1. Install Miniconda (if not already installed):

   • Download from: https://docs.conda.io/en/latest/miniconda.html
   • Run the installer and follow prompts

2. Run the setup script:

   cd ENV_SETUP
   setup_env.bat      # Windows
   # OR
   bash setup_env.sh  # Linux/macOS

3. Activate the environment:

   conda activate EDU

Option B: Manual Setup

1. Create conda environment:

   conda create -n EDU python=3.10 -y
   conda activate EDU

2. Install PyTorch with CUDA:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

3. Install remaining dependencies:

   pip install -r requirements.txt

Verify Installation

python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA: {torch.cuda.is_available()}')"

Expected output:

PyTorch: 2.x.x
CUDA: True

---

📁 Project Structure

segmentation_project/
│
├── train_segmentation.py      # Training script
├── test_segmentation.py       # Inference script
├── failure_analysis.py        # Visualization utilities
├── visualize.py               # Color mapping utilities
├── requirements.txt           # Python dependencies
├── README.md                  # This file
├── HACKATHON_REPORT.md        # Detailed project report
│
├── train/                     # Training data
│   ├── Color_Images/          # RGB training images
│   └── Segmentation/          # Ground truth masks
│
├── Color_Images/              # Validation RGB images
├── Segmentation/              # Validation masks
│
├── data/
│   └── testImages/            # Test images (no ground truth)
│       └── Color_Images/      # RGB test images
│
├── output/                    # Training outputs
│   ├── best_model.pth         # Best model checkpoint
│   ├── config.json            # Training configuration
│   ├── training_curves.png    # Loss/metric plots
│   └── training_report.txt    # Training summary
│
├── failure_analysis/          # Failure case visualizations
│   └── case_*.png             # Sample analysis images
│
└── ENV_SETUP/                 # Environment setup scripts
    └── setup_env.bat          # Windows setup script

---

🚀 Training

Basic Training

conda activate EDU
python train_segmentation.py

Training with Custom Parameters

python train_segmentation.py \
    --train_dir ./train \
    --val_dir . \
    --batch_size 4 \
    --epochs 50 \
    --lr 0.001 \
    --backbone small \
    --output_dir ./output

Training Arguments

Argument	Default	Description
--train_dir	./train	Training data directory
--val_dir	.	Validation data directory
--batch_size	4	Batch size (reduce if OOM)
--epochs	50	Maximum training epochs
--lr	1e-3	Learning rate
--backbone	small	DINOv2 size: small/base/large
--output_dir	./output	Output directory
--resume	—	Resume from checkpoint
--no_amp	False	Disable mixed precision
--no_augment	False	Disable augmentation
--seed	42	Random seed

Training Outputs

After training completes:

• output/best_model.pth — Best model weights
• output/config.json — Training configuration
• output/training_curves.png — Loss and metric plots
• output/training_report.txt — Detailed metrics per epoch

---

🔍 Testing & Inference

Run Inference on Test Images

python test_segmentation.py \
    --model_path ./output/best_model.pth \
    --test_dir ./data/testImages/Color_Images \
    --output_dir ./test_results

Test Arguments

Argument	Default	Description
--model_path	Required	Path to model checkpoint
--test_dir	Required	Directory with test images
--output_dir	./results	Output directory
--visualize	False	Save colored visualizations
--batch_size	1	Inference batch size

Generate Failure Analysis

python failure_analysis.py \
    --model_path ./output/best_model.pth \
    --image_dir ./Color_Images \
    --mask_dir ./Segmentation \
    --output_dir ./failure_analysis \
    --max_images 20

---

📊 Expected Outputs

Training Output

2026-01-23 10:00:00 | INFO | Starting training...
2026-01-23 10:00:05 | INFO | Loaded 2129 training samples
2026-01-23 10:00:06 | INFO | Loaded 1002 validation samples
Epoch 1 [Train]: 100%|██████████| 1064/1064 [02:15<00:00]
Epoch 1 [Val]:   100%|██████████| 501/501 [00:45<00:00]
2026-01-23 10:03:06 | INFO | Epoch 1: Val IoU=0.2374, Val Loss=1.8050
...
2026-01-23 10:45:00 | INFO | Training completed!
2026-01-23 10:45:00 | INFO | Best Val IoU: 0.2769 (Epoch 3)

Test Output

2026-01-23 11:00:00 | INFO | Loading model from ./output/best_model.pth
2026-01-23 11:00:05 | INFO | Processing 100 test images...
2026-01-23 11:01:30 | INFO | Saved predictions to ./test_results/
2026-01-23 11:01:30 | INFO | Test IoU: X.XXXX

---

🔧 Troubleshooting

Common Issues

1. CUDA Out of Memory

RuntimeError: CUDA out of memory

Solution: Reduce batch size:

python train_segmentation.py --batch_size 1

2. Module Not Found

ModuleNotFoundError: No module named 'torch'

Solution: Activate environment and reinstall:

conda activate EDU
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124

3. CUDA Not Available

CUDA: False

Solution:

1. Verify NVIDIA drivers: nvidia-smi
2. Reinstall PyTorch with correct CUDA version
3. Check GPU compatibility

4. Slow Training

Solutions:

• Close unnecessary applications
• Reduce num_workers in config
• Enable mixed precision: --no_amp remove this flag
• Use SSD for data storage

5. Permission Denied

Windows Solution: Run terminal as Administrator

Linux Solution: chmod +x setup_env.sh

---

✅ Hackathon Compliance

Rules Adherence

Rule	Status
✓ Train only on provided dataset	Compliant
✓ No test images used for training	Compliant
✓ Original model architecture	Compliant
✓ Reproducible results	Compliant

Submission Checklist

• train_segmentation.py — Training script
• test_segmentation.py — Inference script
• best_model.pth — Trained model weights
• config.json — Training configuration
• README.md — Setup instructions
• HACKATHON_REPORT.md — Project report
• requirements.txt — Dependencies

---

📝 Citation

If you use this work, please cite:

@misc{offroad_segmentation_2026,
  title={Offroad Semantic Segmentation for Autonomous Navigation},
   author={Rashi innovators},
  year={2026},
  note={Duality AI Code Sprint Hackathon}
}

---

📧 Contact

For questions or issues:

• Discord: Duality AI Community
• Email: [akshaykammar31@gmail.com]

---

Built for the Duality AI Code Sprint Hackathon 2026

Tag @DualityAI on LinkedIn to share your results!