Semantic segmentation model that labels every pixel of desert images into 10 terrain classes. Built for the Duality AI Falcon Hackathon.
| Version | Epochs | Val IoU | Test IoU | Status |
|---|---|---|---|---|
| V1 Baseline | 10 | 0.2972 | 0.2347 | Reference |
| V2 Optimized | 10 | 0.2943 | 0.2347 | Faster training |
| V3 Full Train | 30 | 0.4712 | 0.4700 | Best |
V3 doubles the test IoU compared to V1. Rare classes improved 6 to 48 times.
python -m venv .venv
source .venv/bin/activate
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install numpy pillow matplotlib opencv-python tqdmpython scripts/trainSegmentation.pyOutputs: model weights in models/, training curves in trainStats/.
python scripts/testSegmentation.py --dataDir /path/to/dataset --outputDir predictionsOutputs: evaluation metrics and comparison images in predictions/.
| Component | Details |
|---|---|
| Backbone | DINOv2 ViT-B/14-reg, frozen, self-supervised |
| Head | ConvNeXt-style, 3 blocks, 256 channels |
| Input | 960x540 resized to 546x280 |
| Classes | 11 (background + 10 terrain types) |
| Loss | 0.5 weighted CE + 0.5 Dice |
| Optimizer | AdamW, lr=1e-3, cosine schedule, weight decay 1e-4 |
| Batch Size | 4 |
| Augmentation | Flip, ColorJitter, GaussianBlur |
| Precision | AMP (CUDA autocast) |
| Class | V1 | V2 | V3 |
|---|---|---|---|
| Sky | 0.9449 | 0.9412 | 0.9570 |
| Trees | 0.0907 | 0.0864 | 0.5836 |
| Lush Bushes | 0.0112 | 0.0105 | 0.4771 |
| Dry Grass | 0.1514 | 0.1577 | 0.5372 |
| Dry Bushes | 0.0600 | 0.0675 | 0.3612 |
| Ground Clutter | 0.0491 | 0.0598 | 0.2454 |
| Flowers | - | - | 0.5008 |
| Logs | 0.0356 | 0.0343 | 0.2299 |
| Rocks | 0.2646 | 0.2053 | 0.2431 |
| Landscape | 0.5031 | 0.5485 | 0.4359 |
Offroad_Segmentation_Scripts/
├── docs/ Task description, workflow, hackathon briefs
├── scripts/ trainSegmentation.py, testSegmentation.py, visualize.py
├── models/ Trained model weights (.pth files)
├── experiments/ Baseline and exp1-exp4 outputs
│ ├── baseline/ Original 10-epoch run
│ ├── exp1/ 3 epochs, AdamW
│ ├── exp2/ 5 epochs, cleaner eval
│ ├── exp3/ 10 epochs, SGD
│ └── exp4/ 10 epochs, batch=2, AMP
├── trainStats/ V3 training curves and metrics
├── predictionsBatch1/ V3 batch 1 predictions
├── predictionsBatch2/ V3 batch 2 predictions
├── predictionsBatch3/ V3 batch 3 predictions (best: 0.4700)
├── frontend/ Interactive React report
│ ├── src/pages/ Overview, V1, V2, V3 tabs
│ └── public/ Charts and comparison images
├── envSetup/ Environment setup scripts
└── README.md This file
React + TypeScript dashboard with 4 tabs:
| Tab | Content |
|---|---|
| Overview | V1 vs V2 vs V3 comparison table |
| V1 | Baseline results, training curves, prediction samples |
| V2 | Optimized results, training curves, prediction samples |
| V3 | 30-epoch full training results, per-class breakdown |
Run with:
cd frontend
npm install
npm run devView the full report: SNN_Duality.pdf
- GPU: NVIDIA GeForce RTX 4050 Laptop, 6GB VRAM
- CUDA: 12.1
- PyTorch: 2.5.1