Offroad Semantic Segmentation — Final Submission

Model Summary

Metric	Score
Architecture	DeepLabV3+ with ResNet-101 encoder
Best mIoU (val)	61.93%
Pixel Accuracy (val)	86.10%
Best Epoch	77 / 80 (trained 80 of 150 configured)
Image Size	768 × 768

Per-Class IoU Scores

Class	IoU (%)
Trees	85.51
Lush Bushes	67.93
Dry Grass	69.04
Dry Bushes	48.39
Ground Clutter	39.83
Flowers	46.91
Logs	47.16
Rocks	49.93
Landscape	66.13
Sky	98.48

---

Repository Structure

Final Submission/
├── README.md                  # This file
├── requirements.txt           # Python dependencies (pip)
├── environment.yml            # Conda environment spec
├── configs/
│   ├── deeplab_boost.yaml     # Best performing config (used for final model)
│   ├── deeplab_config.yaml    # Baseline DeepLabV3+ config
│   ├── base_config.yaml       # Base config template
│   └── unet_config.yaml       # Alternate U-Net config
├── src/
│   ├── train.py               # Training script
│   ├── test.py                # Testing & evaluation script
│   ├── model.py               # Model architecture builder
│   ├── dataset.py             # Dataset & augmentation pipeline
│   ├── losses.py              # Combined loss functions (CE + Dice + Focal)
│   ├── metrics.py             # IoU & pixel accuracy metrics
│   ├── utils.py               # Utilities (checkpointing, logging, scheduler)
│   ├── inference.py           # Single-image inference script
│   └── __init__.py
├── data/
│   └── class_mapping.json     # Class ID → index mapping with colors
├── checkpoints/
│   └── best.pth               # Best model weights (epoch 77, 61.93% mIoU)
└── outputs/
    ├── iou_scores.txt         # Detailed per-class IoU scores
    ├── confusion_matrix.png   # Normalized confusion matrix
    ├── loss_curve.png         # Training/validation loss over epochs
    ├── miou_curve.png         # mIoU progression over epochs
    ├── predictions/           # Test set predictions (segmentation masks)
    └── visualizations/        # Side-by-side visual comparisons

---

Environment Setup

Option A: pip (recommended)

python -m venv venv
# Windows:
venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate

pip install -r requirements.txt

Option B: Conda

conda env create -f environment.yml
conda activate segmentation

Requirements

• Python 3.10+
• CUDA-capable GPU (6GB+ VRAM recommended)
• PyTorch 2.0+ with CUDA support

---

How to Reproduce Results

1. Place the Dataset

Download the competition dataset and place it so the directory structure is:

data/
├── train/
│   ├── Color_Images/    # Training RGB images
│   └── Segmentation/    # Training ground-truth masks
├── val/
│   ├── Color_Images/    # Validation RGB images
│   └── Segmentation/    # Validation ground-truth masks
├── test/
│   └── Colour_Images/   # Test RGB images (no masks)
└── class_mapping.json

2. Train the Model

python src/train.py --config configs/deeplab_boost.yaml

To resume from a checkpoint:

python src/train.py --config configs/deeplab_boost.yaml --resume checkpoints/best.pth

3. Evaluate on Validation Set (reproduce 61.93% mIoU)

python src/test.py --config configs/deeplab_boost.yaml --checkpoint checkpoints/best.pth --eval_val

4. Generate Test Predictions

python src/test.py --config configs/deeplab_boost.yaml --checkpoint checkpoints/best.pth

Predictions are saved to outputs/predictions/.

5. Optional Enhancements

Test-Time Augmentation (TTA) — multi-scale + flip averaging for ~0.5% mIoU boost:

python src/test.py --config configs/deeplab_boost.yaml --checkpoint checkpoints/best.pth --eval_val --tta

CRF Post-Processing — edge-aware refinement for sharper boundaries:

python src/test.py --config configs/deeplab_boost.yaml --checkpoint checkpoints/best.pth --eval_val --crf

---

Expected Outputs

• Evaluation: Prints mIoU, pixel accuracy, and per-class IoU to console. Saves scores to outputs/iou_scores.txt and confusion matrix to outputs/confusion_matrix.png.
• Test inference: Saves predicted segmentation masks as PNG files in outputs/predictions/ and visual overlays in outputs/visualizations/.

---

Key Configuration (deeplab_boost.yaml)

Parameter	Value	Purpose
Image size	768×768	Larger crops capture more context
Batch size	3 (×4 accum = eff 12)	Fits in 6GB VRAM with grad accumulation
Encoder	ResNet-101 (ImageNet pretrained)	Strong feature extractor
Optimizer	AdamW (lr=3e-4, wd=0.01)	Stable convergence
Scheduler	OneCycleLR (cos anneal)	Warm-up + aggressive LR decay
Loss	CE(0.4) + Dice(0.4) + Focal(0.2)	Balanced per-pixel + per-class learning
Augmentations	Flip, rotate(45°), scale(0.4), elastic, color jitter, random crop	Robustness to desert variation
Mixed precision	Enabled	2× faster training, lower VRAM
Early stopping	20 epochs patience	Prevents overfitting