IEEE MLSP Data Competition

The Sampling-Assisted Pathloss Radio Map Prediction Data Competition, IEEE MLSP 2025

Motivation

Accurate prediction of pathloss (PL) maps in indoor environments is fundamental to applications including fingerprint-based localization, user-cell site association and path planning.

Traditional methods such as ray-tracing are usually computationally expensive and time-consuming. There exist gap between simulated and realistic data and lack generalization ability to different, unseen environments.

Contributions

Thus, we developed a runtime-efficient U-Net model for 2D indoor pathloss predition under sparse sampling conditions, integrating environment-aware geometric features to enhance cross-environment generalization. Our method is data-driven, utilizing lighter and more efficient network and real-world measurements, and adaptive to different environments.

Our main contributions are twofold:

  • Integrated environment-aware geometric features such as accumulated transmittance, obstruction counts to enhance cross-environment robustness
  • Designed a lightweight U-Net model to maximize feature learning and generalization, while minimize runtime
The architecture we used for the competition is a U-Net with extracted features.

Results

We achieved average total runtime of 14.36 ms and received 2nd place in this data competition.

The ablation studies shows 2 important insights:

  • Setting 6 which integrates the full set of engineered inputs achieves the lowest validation RMSE.
  • Even sparse PL inputs provide essential guidance for accurate map reconstruction.
The ablation studies we did (R: reflectance; T: transmittance; D: distance; Sampled GT: 0.5% sampled ground truth path loss; log-D: log-scaled distance; OC: obstruction count map; Tsum: accumulated transmittance; FSPL: freespace path loss; T'sum: Tsum+FSPL)

For more system infomation please refer to our paper.