IPP-Net: A Generalizable Deep Neural Network Model for Indoor Pathloss Radio Map Prediction

TL;DR

This work targets indoor pathloss radio map prediction, where indoor obstructions and frequency variation challenge outdoor-tuned approaches. It introduces IPP-Net, a UNet-based DNN that ingests five-channel inputs, including a model channel built from a modified 3GPP InH model with $L_{i,j}$ and $\Delta_{i,j}$ and a frequency indicator, and employs curriculum learning across three tasks to improve generalization. Trained on ray-tracing data, IPP-Net achieves a weighted RMSE of $9.501$ dB across three tasks and ranks second in the ICASSP 2025 First Indoor Pathloss Radio Map Prediction Challenge, while delivering fast inference (≈10 ms per radio map). The approach demonstrates a practical path for accurate, scalable indoor radio map prediction, combining indoor priors with deep representation learning to support planning and optimization.

Abstract

In this paper, we propose a generalizable deep neural network model for indoor pathloss radio map prediction (termed as IPP-Net). IPP-Net is based on a UNet architecture and learned from both large-scale ray tracing simulation data and a modified 3GPP indoor hotspot model. The performance of IPP-Net is evaluated in the First Indoor Pathloss Radio Map Prediction Challenge in ICASSP 2025. The evaluation results show that IPP-Net achieves a weighted root mean square error of 9.501 dB on three competition tasks and obtains the second overall ranking.

Figures (1)

• Figure 1: An example radio map predicted by IPP-Net