Scaling Cheap Synthetic Pretraining Improves Indoor Radio-Map Prediction
We address both problems by introducing a generator that automatically creates diverse indoor layouts and a fast, open-source approximate ray tracer that computes radio maps for them.
ProofPaper ↗
Key points
- Real measurements are too limited to train an accurate indoor pathloss predictor that works across a wide range of buildings.
- Large-scale synthetic pretraining offers a way forward, but real building layouts are also scarce, and accurate ray-tracing simulators are slow, expensive, and often proprietary.
- Despite simulator mismatch, our ResNet-101 model reduces error by 15% on average relative to the best known methods across five indoor radio-map prediction tasks and improves transfer on a small real-measurement dataset.
- These results show that pretraining on large amounts of cheap simulated data can improve indoor radio-map prediction, even when real measurements and high-quality simulations are limited.
Sources (1)
- [1]Scaling Cheap Synthetic Pretraining Improves Indoor Radio-Map PredictionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 06:09 PM
We address both problems by introducing a generator that automatically creates diverse indoor layouts and a fast, open-source approximate ray tracer that computes radio maps for them.
Real measurements are too limited to train an accurate indoor pathloss predictor that works across a wide range of buildings.
Extractive summary: sentences quoted from the sources.