Offroad Semantic Segmentation Model A semantic segmentation system for detecting vegetation, objects, and terrain features in offroad/desert environments. Built with DINOv2 backbone and a ConvNeXt-style segmentation head. Features

• Vegetation Detection: Trees, Lush Bushes, Dry Grass, Dry Bushes
• Object Detection: Rocks, Logs, Ground Clutter
• Terrain Analysis: Landscape and Sky segmentation
• Pixel-level Coverage Metrics: Class distribution and percentages

Datasets used: https://falcon.duality.ai/secure/documentation/hackathon-segmentation-desert?utm_source=hackathon&utm_medium=instructions&utm_campaign=codesprint

Project Structure desert/ ├── Offroad_Segmentation_Scripts/ │ ├── train_segmentation.py # Training │ ├── test_segmentation.py # Validation │ ├── visualize.py # Visualization │ ├── ENV_SETUP/ # Environment setup │ └── train_stats/ # Training metrics ├── inference_pipeline.py # Inference pipeline ├── Offroad_Segmentation_Training_Dataset/ │ ├── train/ # Training data │ └── val/ # Validation data └── README.md

Quick Start

1. Setup Environment cd Offroad_Segmentation_Scripts/ENV_SETUP/ pip install torch torchvision torchaudio pip install pytorch-lightning torchmetrics pillow opencv-python numpy matplotlib tqdm

2. Train Model cd Offroad_Segmentation_Scripts/ python train_segmentation.py

3. Run Inference python inference_pipeline.py
   --image_dir /path/to/images
   --model_path Offroad_Segmentation_Scripts/segmentation_head.pth
   --output_dir ./predictions

Outputs

• masks/ → Raw segmentation masks (class IDs 0–9)
• masks_color/ → Colored masks for visualization
• visualizations/ → Side-by-side input + prediction
• analysis_summary.json → Per-image statistics
• class_legend.png → Color legend