📶 This project aims to enhance Radio Maps with Super Resolution and is inspired by DLSS. The project is part of the course DD2430 at KTH and is done in collaboration with Ericsson.
Left: Example of a radio map data sample with transmitter (blue). Right: Example of a radio map over San Francisco with transmitter (red).
The table below show results from upsampling the low resolution maps using bicubic interpolation and the trained super resolution model (PAN).
| Metric | Bicubic | PAN |
|---|---|---|
| PSNR | 21.97 | 24.21 |
| SSIM | 0.289 | 0.377 |
| RMSE | 24.21 | 18.69 |
- Install uv by running
curl -LsSf https://astral.sh/uv/install.sh | sh. - Sync environment:
uv sync - Run commands:
uv run <command> - (Dev): Install pre-commit hooks:
pre-commit install
- Might need
LLVMinstalled...
See src/poc/readme.md for detailed instructions on how to run the proof of concept code. Available commands:
uv run generate: Generate synthetic radio map data.uv run train: Train the super-resolution model.uv run test: Evaluate the model.
This codebase is developed and tested on Linux systems and WSL2.
Using an NVIDIA GPU for training is recommended. NVIDIA T4 and L4 have been tested.