LoveDA Semantic Segmentation Benchmark

Project Documentation

For architectural overview and component relationships, see ARCHITECTURE.md
Detailed technical specifications and implementation notes available in DOCS.md

High-Level Objectives

• Build benchmarking scripts for LoveDA dataset
• [✔] Implement comprehensive comparison of EO models (in testing)
• [✔] Finalize visualization system (v1 implemented)

Mid-Level Progress

✅ Completed:

• Model initialization & configuration (torchvision models)
• PyTorch Lightning integration
• CLI benchmarking interface
• Automated documentation updates

✅ Completed:

• Dataset ingestion & transformation pipeline
• Class weight integration

🔄 In Progress:

• CI/CD pipeline integration

Current Implementation Status

Latest Benchmarks
See results/README.md for latest metrics

Recent Changes

• Added UNet++ and PSPNet architectures
• Integrated mixed precision training support
• Automated benchmark report comparison system
• CI/CD readiness checks implemented

Completed in Testing Phase

1. Data augmentation pipeline (v2.1 implemented)
2. Prediction visualization hooks (see /src/lovebench/visualization.py)
3. Extended metrics suite (Dice, IoU, Precision/Recall)

Implementation Questions

✅ Resolved:

1. Dataset split ratios: 70-15-15 train/val/test
2. Class weights calculated globally
3. Cloud detection handled at dataset level
4. CUDA visibility confirmed through environment validation checks
5. CLI working directory standardized to project root

Usage

Environment: Torchgeo_Benchmarks (Python 3.8)
Working Directory: Execute from project root (/home/colm-the-conjurer/VSCode/workspace/LoveDA-Benchmarking)

# Run full benchmark suite
python -m lovebench.cli \
  --models deeplabv3 fcn_resnet50 lraspp_mobilenet unetplusplus pspnet \
  --data-root ./data/LoveDA \
  --output benchmarks

Dependencies

• Python 3.8+
• PyTorch 2.0.1+
• TorchGeo 0.5.2+
• PyTorch Lightning 2.1.0+
• CUDA 11.8+