PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Cheap Synthetic Pretraining Improves Indoor Radio-Map Prediction

Khoren Petrosyan, Artashes Mkrtchyan, Rafayel Mkrtchyan, Hrant Khachatrian, Theofanis P. Raptis

arXiv:2610.19015v1cs.LGcs.AI

TL;DR

Indoor radio-map prediction lacks abundant high-quality measurements, layouts, and affordable accurate simulation. The paper generates synthetic layouts and approximate radio maps for large-scale pretraining, then fine-tunes on benchmarks. ResNet101 improves across five public-leaderboard tasks, while both backbones show predictable aggregate scaling with non-monotone later checkpoints.

  • Problem

    High-quality indoor radio-map data are limited, while accurate ray-tracing simulators are slow, expensive, and often proprietary.

    Source: §1 · p. 1

  • Method

    The paper procedurally generates diverse layouts and approximate radio-map labels, pretrains encoder-decoders on 128M samples, and fine-tunes them on high-quality benchmarks.

    Source: §1 · p. 2

  • Results

    ResNet101 improves on all five public-leaderboard tasks, and 2M–16M power-law fits predict normalized downstream RMSE at 32M–128M with 3.33% MAPE for ResNet101 and 7.10% for Swin-T.

    Source: §6 · p. 7, Conclusion · p. 8

  • Takeaways & Limitations

    Cheap approximate simulation can make scarce high-quality indoor radio-map data more useful, but high-fidelity simulation and real measurements remain necessary for adaptation and validation.

    Source: Conclusion · p. 8

    The small real dataset establishes calibrated predictive information, not deployment readiness or end-to-end transfer.

    Source: Discussion and limitations · p. 8

Abstract

from arXiv · show

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. 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. Using this pipeline, we generate 128M synthetic samples, pretrain ResNet-101 and Swin-T encoder-decoders, and fine-tune them on high-quality benchmark data. 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. Both architectures show predictable data-scaling trends: power laws fitted on 2M-16M samples predict the mean normalized downstream RMSE at 32M-128M with 3.33% mean absolute percentage error for ResNet-101 and 7.10% for Swin-T. 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.

1 Introduction

Indoor radio-map prediction has scarce high-quality data, while accurate ray tracing is slow and often proprietary. The paper tests scalable pretraining using procedurally generated layouts and cheap approximate labels, then evaluates transfer and scaling.

1 Introduction

  • The approach scales pretraining rather than proposing a new architecture or more accurate propagation model.The generator omits diffraction, scattering, and 3D propagation, leaving scale and diversity to be tested against simulator mismatch.Source: §1 · p. 2
  • A 128M-sample synthetic corpus and configurable input–label pipeline address the shortage of layouts and scalable labels.The corpus occupies roughly 6 TB; generation code is public and the dataset will be released.Source: §1 · p. 2
  • The study compares 10 encoder-decoders from four families under a fixed-FLOP protocol, selecting ResNet101 and Swin-T for scaling runs.It also tests frozen-feature transfer on real RSSI and probes propagation physics and geometric symmetries.Source: §1 · p. 2
  • Downstream transfer improves predictably with cheap synthetic pretraining scale: a data-only power law fitted on 2M–16M checkpoints predicts 2M-normalized mean RMSE at larger scales with low extrapolation error for both models.Source: §1 · p. 2
  • ResNet101 improves on the best public-leaderboard references across five ICASSP/MLSP tasks.The paper also examines simulator fidelity and missing material information.Source: §1 · p. 2

2 Related Work

Prior radio-map methods use image-to-image networks and often add propagation-specific features or architectures. This work situates cheap synthetic pretraining as a scalable alternative, with downstream fine-tuning aligning approximate simulation to high-quality data.

2 Related Work

  • Radio-map systems have progressed from RadioUNet’s image-to-image estimation to stronger backbones, sparse inputs, and physics-informed features.Recent examples include SIP2Net, IPP-Net, and sparse-measurement features.Source: §2 · p. 2
  • Best-performing scarce-data systems often rely on bespoke propagation features, feature engineering, or benchmark-specific architectures.The paper asks whether scaling cheap synthetic pretraining can reduce this dependence while improving transfer across tasks.Source: §2 · p. 2
  • Approximate-simulator pretraining uses a simplified physics engine for scalable supervision, then fine-tunes on high-quality simulated and measured data.The ICASSP and MLSP benchmarks instead provide fixed datasets of high-fidelity ray-traced maps.Source: §2 · p. 2
  • The scaling analysis fixes the model family after architecture selection and predicts downstream performance after fine-tuning, not pretraining loss alone.Its scaling variable is synthetic pretraining data or effective sample exposure.Source: §2 · p. 3

3 Indoor Radio-Map Prediction Tasks

The prediction task maps image-like floor-plan and transmitter inputs to dense pathloss maps evaluated in dB. Five benchmark settings test generalization across layouts, frequencies, and antenna patterns, as well as sparse-measurement reconstruction.

3 Indoor Radio-Map Prediction Tasks

  • Each sample combines rasterized environment and transmitter information into an input tensor, and the model predicts a dense pathloss map.Evaluation uses RMSE in dB on held-out maps.Source: §3 · p. 3
  • Five settings comprise three ICASSP tasks for unseen layouts, carrier frequencies, or antenna patterns and two MLSP sparse-sampling rates.The sparse sampling rates are 0.02% and 0.5% of map pixels.Source: §3 · p. 3
  • A shared 11-channel featurizer is used for synthetic pretraining and downstream fine-tuning.Channels encode wall reflectance and transmittance, transmitter distance and antenna gain, frequency, valid pixels, floor plan, and sparse measurements.Source: §3 · p. 3

4 Scalable Synthetic Radio-Map Generation

A procedural floor-plan rasterizer and accelerated 2D ray tracer generate synthetic inputs and pathloss labels. The corpus is large and separated from benchmark layouts, but the tracer differs materially from high-fidelity simulation and omits several propagation effects.

4 Scalable Synthetic Radio-Map Generation

  • A procedural rasterizer supplies floor-plan inputs and a Numba-accelerated 2D ray tracer supplies pathloss targets.It samples building geometry and propagation parameters; the tracer launches 46,080 rays, with up to five reflections and ten wall transmissions.Source: §4 · pp. 3–4
  • The tracer omits diffraction, scattering, and 3D propagation, and replaying benchmark inputs yields 10.91 dB median per-map RMSE against Ranplan.The generator uses no benchmark layouts or labels and its distributions were not fitted to them.Source: §4 · p. 4
  • A finite nearest-neighbor audit found no exact matches, but it does not exhaustively check the 128M-sample corpus.The audit compared 25 ICASSP layouts with 50,000 independently generated layouts.Source: §4 · p. 4
  • The generator creates 128M unique synthetic samples, requiring approximately 6 TB of storage.Training checkpoints span 2M–128M unique samples and 192M–256M effective exposures.Source: §4 · p. 4

5 Experiments

The experiments select architectures under a fixed-compute protocol, then pretrain selected models at multiple synthetic-data exposures and fine-tune each checkpoint on downstream tasks. Figure 2 compares task-level and normalized downstream RMSE across scaling budgets.

5 Experiments

  • Ten encoder-U-Net variants from four families are compared under a fixed-step, fixed-FLOP protocol, with ResNet101 selected for the main scaling runs.The search uses 350,000 steps and approximately 6 × 10^18 theoretical training FLOPs; each variant is selected by final training loss.Source: §5 · p. 4, Table 2 · p. 5
  • ResNet101 is pretrained on the full synthetic corpus, with decayed checkpoints saved at exposures from 2M to 256M samples.The run uses a sample-based warmup-stable-decay schedule, with each target checkpoint forked at 0.9 of its target exposure.Source: §5 · p. 5
  • Swin-T is also pretrained at seven scales from 2M to 128M and fine-tuned on all five tasks, using one pretraining seed per backbone.The results do not estimate variability across independent pretraining runs.Source: §5 · p. 5
  • Each checkpoint is fine-tuned separately on every task using a shared peak learning rate selected on ICASSP Task 3 rather than tuned per task.The selected peak learning rate is 5 × 10^-4, and checkpoints are selected by held-out internal-validation RMSE.Source: §5 · p. 5
  • Figure 2 shows task-level public-leaderboard RMSE and each backbone’s mean RMSE normalized by its own 2M result across scaling budgets.The dotted divider marks the 2M–16M fitting range; Swin-T’s lower normalized RMSE mainly reflects higher 2M errors and larger relative improvement, while 128M mean RMSEs are fairly similar.Source: Figure 2 · p. 6

6 Results

The results assess benchmark transfer, aggregate downstream scaling, and the effects of simulator fidelity and missing material inputs. ResNet101’s selected-scale benchmark comparison and the scaling-law results are qualified by differences across tasks and by input-information constraints.

6 Results

  • The first-epoch ResNet101 benchmark comparison reports the best public RMSE across 2M–128M scales after internal-validation epoch selection, rather than one common checkpoint.The comparisons use named public-leaderboard entries; ICASSP tasks vary buildings, frequencies, and antennas, while MLSP provides sparse target-domain pathloss.Source: §6 · p. 6, Table 3 · p. 7
  • 3.33% MAPE for ResNet101 and 7.10% for Swin-T quantify predictions of held-out 32M–128M normalized mean RMSE from fits using 2M–16M checkpoints.The aggregate fits are not guarantees for individual tasks; improvement is not strictly monotone, and the ResNet101 fit has 4.00% error through 256M, including repeated data.Source: §6 · p. 7
  • With 16M pretraining, frozen five-task mean RMSE is 10.50/9.02/8.99 dB for zero/two/five-reflection simulators, versus 6.79/6.60/6.58 dB after full fine-tuning.In scarce-label experiments, gains reach 19.3% on ICASSP T2 and 14.8% on MLSP 0.02; missing material channels also reduce the 0.5%-support 8M–64M mean per-building gain from 0.252 to 0.121 dB.Source: §6 · p. 7

7 Real-Measurement Transfer

The real-measurement experiment tests whether frozen synthetic-pretrained pathloss predictions add information to a calibrated RSSI regressor. The combined model improves over a geometric baseline at 8M, while the limited number of AP-location groups leaves later scale ordering unresolved.

7 Real-Measurement Transfer

  • At 8M, adding the frozen prediction lowers RMSE from 5.878 to 5.117 dB (12.9%) across leave-one-AP-location-out folds.Only the combined model significantly improves on the fixed-feature baseline; the supported conclusion is incremental information beyond that geometric baseline.Source: §7 · p. 7
  • The combined model’s RMSE is 5.117, 5.323, and 5.234 dB at 8M, 128M, and 256M, but comparisons with 8M do not establish later ordering.The study uses 16 independent AP-location groups, and neither continued real-data improvement nor degradation is established.Source: §7 · p. 7
  • Frozen MLSP models show diminishing returns around 8M while fine-tuned transfer improves later, but lack of target adaptation is only a plausible explanation.End-to-end testing needs larger, more diverse measured datasets.Source: §7 · p. 8

8 What Does the Model Learn?

Controlled probes compare the 128M ResNet101 checkpoint with physical invariants and closed-form expectations in synthetic diagnostic scenes. The probes show several forms of physically meaningful structure beyond benchmark-specific interpolation.

8 What Does the Model Learn?

  • The 128M ResNet101 checkpoint reproduces free-space frequency scaling with 0.45% MAPE and distance scaling with 1.10% MAPE.It also calibrates single-wall loss with mean median absolute error 0.41 dB across 3–20 dB wall losses.Source: §8 · p. 8
  • The model shows D4 scene-isometry consistency with 0.51 dB median absolute difference after alignment and open-field angular isotropy with 0.037 dB median angular residual.Source/receiver reciprocity over representative chords has 0.70 dB MAE.Source: §8 · p. 8
  • The probes are not tuned to the benchmark competitions or real-measurement regression task, supporting the authors’ view that learned invariances reduce the likelihood of evaluation-specific adaptation.The authors state this increases confidence that the representation can be useful in other indoor radio-map settings where labeled measurements or high-fidelity simulations are scarce.Source: §8 · p. 8

Discussion and limitations

The proposed workflow uses broad procedural pretraining, then adapts with trusted simulations or measurements; its evidence is limited to indoor radio maps. Several modeling and evaluation boundaries remain open.

Discussion and limitations

  • Evidence is limited to indoor radio maps, even though the recipe may apply to other physics-governed spatial tasks.A compute-optimal allocation between accurate and approximate labels remains open; the fidelity study holds sample count fixed, not generation compute.Source: Discussion and limitations · p. 8
  • The simulator omits diffraction, scattering, and 3D propagation, and the reflection-depth study does not establish robustness to those effects or arbitrary floor-plan generators.The study also uses data-only scaling with one pretraining seed per backbone and does not jointly model parameters and compute.Source: Discussion and limitations · p. 8
  • The small real dataset establishes calibrated predictive information, not deployment readiness or end-to-end transfer.Self-supervised, masked, foundation-model, domain-adaptation, and generative-pretraining baselines remain open comparisons.Source: Discussion and limitations · p. 8

Conclusion

The paper concludes that cheap approximate simulation can make scarce high-quality indoor radio-map data more useful. Its generator builds varied floor plans and paired inputs and dense pathloss targets for pretraining and fine-tuning.

Conclusion

  • Cheap approximate simulation can make scarce high-quality indoor radio-map data more useful, although fidelity and material information still matter.ResNet101 improves all five public-leaderboard tasks; both backbones show predictable aggregate scaling with non-monotone later checkpoints.Source: Conclusion · p. 8

A Synthetic Input-Output Generation

  • The generator creates variable-width floor plans by arranging cores, corridors, rooms, and openings, then recursively partitioning remaining interior space.Room sizes vary within a plan through a spatially varying heavy-tailed area threshold.Source: A Synthetic Input-Output Generation · p. 12
  • Material maps assign wall reflectance and transmittance losses, while transmitter positions and carrier frequencies are sampled for each scene.Wall-pixel normals support the reflection model; the frequency range is 400–10,000 MHz.Source: A Synthetic Input-Output Generation · p. 12
  • Antenna patterns are sampled independently of floor plans and include isotropic or varied directional patterns.Patterns are stored as both a 360-sample array and a function description.Source: A Synthetic Input-Output Generation · p. 12
  • The ray tracer combines antenna, free-space, transmission, and reflection losses to produce dense pathloss targets, with line-of-sight backfill for pixels missed by reflected rays.The simulator omits diffraction and scattering, and stores the minimum pathloss assigned to each pixel.Source: A Synthetic Input-Output Generation · pp. 12–13
  • Figure 3 pairs reflectance, transmittance, and radiation-pattern inputs with a floor plan and dense target, using the same channel semantics for pretraining and fine-tuning.The paired samples are generated without an external CAD model.Source: Figure 3 · p. 12, A Synthetic Input-Output Generation · p. 12

B Benchmark separation audit

The benchmark and transfer studies assess layout separation, scaling, simulator fidelity, scarce-label learning, real-data calibration, and controlled physics behavior. Results show useful transfer despite approximation, alongside limits from fidelity, materials, and evaluation scope.

B Benchmark separation audit

  • The generated layouts and material pairs differ broadly from benchmark distributions, and the finite layout audit found no exact benchmark–synthetic matches.Benchmark layouts had median nearest-neighbor Chamfer distance 1.451% versus 0.624% for synthetic queries, and median IoU 0.263 versus 0.549; the audit cannot rule out overlap across all 128M samples.Source: B Benchmark separation audit · p. 13

C Transfer across backbones and pretraining scales

  • First-epoch ResNet101 scores reduce scratch error by 21.39%, 20.28%, 27.06%, 48.39%, and 45.46% across ICASSP T1/T2/T3 and MLSP 0.02/0.5, respectively.At 128M versus 2M pretraining samples, the corresponding reductions are 13.77%, 1.88%, 29.50%, 22.15%, and 16.07%.Source: C Transfer across backbones and… · p. 13
  • A data-only power law fitted on 2M–16M predicts ResNet101’s normalized mean RMSE at 32M–128M with 3.33% MAPE, while task curves remain variable and non-monotonic.Fits use 32M–128M as held-out scales; one pretraining run per backbone limits conclusions about seed variability.Source: C Transfer across backbones and… · p. 13, Figure 4 · p. 15

D Reflection-depth ablation at 16M samples

  • At fixed 16M samples, reflection-depth comparisons distinguish direct simulator fidelity from learned transfer; full fine-tuning substantially narrows the fidelity gap.The study does not establish an optimum at fixed generation compute or validate omitted diffraction, scattering, or 3D effects.Source: Table 5 · p. 16, D Reflection-depth ablation at 16M… · p. 16
  • With scarce labels, the largest reported reductions from depth zero are 19.3% for ICASSP T2 with 60 maps and 14.8% for MLSP 0.02 with 120 maps.Higher fidelity is not strictly better in every cell; depth-zero and depth-five results tie at 1,500 maps for ICASSP at displayed precision.Source: D Reflection-depth ablation at 16M… · p. 16, Table 6 · p. 16

E Material-channel removal and a failure case

  • At 0.5% support, 8M–64M improvement averages 0.252 dB per building with material channels and 0.121 dB without them.Missing materials attenuate gains in this setting, but do not establish the cause of real-data saturation because the real experiment uses RSSI calibration without sparse supports.Source: E Material-channel removal and a… · p. 17

F Real-data calibration and uncertainty

  • At 8M, combined frozen-feature real-RSSI calibration reaches RMSE 5.117 dB, versus 5.878 dB for fixed features alone; prediction-only does not establish significant superiority to fixed features.Comparisons with 128M and 256M do not resolve their ordering relative to the best numerical combined RMSE at 8M.Source: F Real-data calibration and uncertainty · p. 17, Figure 5 · p. 18

G Additional Physics Probes

  • Across controlled probes, pretrained predictions follow free-space distance and frequency trends more closely and show smaller rotational, wall-shadowing, and spatial-map errors than the Task 3 baseline.For the 12 dB wall probe, predicted blocked-minus-unblocked contrast is 11.57 dB for the pretrained model and 8.71 dB for the baseline; four exact sparse supports reduce pretrained-model RMSE from 0.40 to 0.19 dB.Source: G Additional Physics Probes · pp. 19, 20, 21, 22, 24
Loading 2610.19015v1…