PaperlayerView exact revision on arXiv

Source-linked AI summary

RadioVIL: Anomaly-Aware Diffusion Models for Radio Map Inpainting and Zero-Shot Vehicle Localization

Ruixin Zhao, Xiucheng Wang, Qiming Zhang, Nan Cheng, Ruijin Sun, Conghao Zhou

arXiv:2608.16167v1eess.SPcs.LG

TL;DR

Sparse radio measurements make high-fidelity map reconstruction and hidden-vehicle detection ill-posed, while pixel-level methods erase scattering signatures. RadioVIL combines diffusion priors with anomaly-aware layer-wise optimization, achieving the best LPIPS of 0.0587 and 75.20% vehicle recall with 3.31-meter average error.

  • Problem

    Sparse measurements make radio-map reconstruction and dynamic-target detection ill-posed, while pixel-level completion erases high-frequency vehicle-scattering signatures.

  • Method

    RadioVIL combines a frozen DDPM structural prior with layer-wise DMILO and L1-regularized sparse deviations to isolate vehicle anomalies for zero-shot localization.

  • Results

    0.0587 LPIPS was the best inpainting score, while zero-shot localization achieved 75.20% recall and 3.31-meter average error under sparse masked observations.

  • Takeaways & Limitations

    Sparse wireless measurements can support zero-shot hidden-vehicle localization when guided by structural generative priors and anomaly-aware inverse optimization.

Abstract

from arXiv · show

High-precision radio map construction is essential for emerging 6G Integrated Sensing and Communication (ISAC) applications, including digital twins and intelligent transportation. However, existing deep learning methods predominantly treat this as a pure image completion task, resulting in over-smoothed reconstructions that fundamentally erase high-frequency scattering signatures of dynamic physical entities such as hidden vehicles. To overcome this, we propose RadioVIL, an efficient two-stage framework that reformulates joint radio map inpainting and zero-shot vehicle localization as a prior-guided physical inverse problem. Specifically, we first train a Denoising Diffusion Probabilistic Model (DDPM) to capture the structural generative prior of the environment. During inference from highly sparse measurements, we employ a Diffusion-based Mediating Intermediate Layer Optimization (DMILO) algorithm. By optimizing an L1-regularized sparse deviation term, DMILO mathematically isolates vehicle scattering anomalies layer-by-layer without unfolding the entire denoising chain. Extensive experiments demonstrate that while conventional reconstruction baselines fail to detect hidden vehicles, and the zero-shot diffusion baseline achieves only limited detection ability due to forced semantic harmonization, RadioVIL preserves authentic physical textures, yielding the best LPIPS of 0.0587 in our evaluation. Uniquely, it unlocks accurate zero-shot vehicle localization directly from sparse radio maps, securing a 75.20% Recall and a 3.31-meter average error, paving a robust way for ISAC at the 6G edge.

I. INTRODUCTION · II. PRELIMINARIES AND SYSTEM MODEL

The paper frames radio-map construction as a physically grounded challenge for 6G ISAC, where sparse sensing and occlusions hinder faithful environmental representation and hidden-vehicle detection. RadioVIL addresses this through a DDPM prior and anomaly-aware DMILO optimization that preserves scattering signatures during inpainting and localization.

  • I. INTRODUCTION: Radio maps connect environmental geometry and materials with spatial signal strength, supporting digital twins, autonomous driving, path planning, and dynamic vehicle localization.
  • I. INTRODUCTION: Crowdsourced sensing is constrained by deployment costs, physical occlusions, and limited access to drivable road regions.
  • I. INTRODUCTION: Pixel-level completion methods such as RadioUNet and RME-GAN favor smooth reconstructions that erase local variations and vehicle scattering footprints.Mean Squared Error supports macroscopic path-loss trends but hinders hidden-target sensing by suppressing high-frequency structure.
  • I. INTRODUCTION: DDPMs capture complex data manifolds and have been extended to dynamic maps, physics-informed modeling, 3D synthesis, and accelerated inference.RadioMamba addresses the computational burden of high-resolution feature extraction.
  • I. INTRODUCTION: Extreme sparsity makes joint radio-map inpainting and sensing difficult because RePaint semantically harmonizes masked regions with the background.This blending suppresses hidden-vehicle anomalies that contradict a smooth structural prior.
  • I. INTRODUCTION: Applying physical constraints across the entire denoising graph incurs high memory cost and may disrupt the generative manifold.
  • I. INTRODUCTION: RadioVIL reformulates inpainting and zero-shot vehicle localization as an anomaly-aware, prior-guided physical inverse problem.Its two-stage design dynamically injects hard physical constraints instead of relying on end-to-end hallucination or rigid pixel replacement.
  • I. INTRODUCTION: The DDPM–DMILO framework uses layer-wise L1-regularized sparse deviation optimization to isolate high-frequency scattering anomalies and preserve hidden-target electromagnetic signatures.This strategy is designed to overcome the semantic over-smoothing flaw of RePaint and traditional deep learning methods.

A. Denoising Diffusion Probabilistic Models

DDPMs model wireless environments by progressively corrupting clean data and learning to reverse that process. Once trained, the frozen denoising network captures the physical environment’s structural prior as a generative dictionary of plausible data.

  • Forward diffusion: DDPMs progressively corrupt clean samples into a sequence of noisy latent variables and support wireless environment reconstruction.They are presented as state-of-the-art generative frameworks for data synthesis and wireless environment reconstruction.
  • Reverse diffusion: A U-Net noise predictor is trained to estimate injected Gaussian noise at each timestep, reversing the forward diffusion process.Training minimizes a variational lower bound objective.
  • Learned prior: The converged frozen denoising network encapsulates the physical environment’s structural prior and represents its physically plausible data manifold.It functions as a generative dictionary for the environment.

B. System Model and Problem Formulation

The system models radio-map recovery on a discretized grid with a binary mask defining sparse, physically accessible measurements while hiding vehicle locations. It decomposes the complete map into smooth background and sparse vehicle scattering, enabling hidden-vehicle inference from residual perturbations and joint reconstruction-localization.

  • System Model: The physical space is discretized as an H × W grid, with the complete radio map R∗∈R^H×W containing environmental information, including hidden vehicles.The formulation also introduces a binary physical mask M ∈ {0, 1}^H×W to represent sensing capability.
  • Sensing Mask: Observable regions comprise buildings, base stations, narrow road-edge bands, and 5%–10% random free-space samples, while roads and potential vehicle regions remain masked.This creates a zero-knowledge zone for vehicle locations.
  • Radio-Map Decomposition: The complete radio map is decomposed into a smooth background and sparse vehicle-induced scattering footprints represented through an effective scattering footprint operator K(·).The sparse vehicle component captures scattering sources caused by hidden vehicles.
  • Indirect Vehicle Inference: Although vehicle pixels are masked, their scattering footprints perturb observable samples, so hidden vehicles are inferred from residuals where M(p) = 1 rather than direct observations.This preserves the strict zero-knowledge assumption while retaining indirect physical observability.
  • Problem Formulation: Given sparse measurements y and a pre-trained diffusion prior, the objective is to recover R∗ and extract hidden vehicle coordinates.The setting links the propagation environment, accessible sensing domain, sparse observations, and dense radio-map reconstruction target.

III. PROPOSED RADIOVIL FRAMEWORK · A. Prior Adaptation and Layer-wise Deterministic Mapping

RadioVIL is a constrained bilevel pipeline that combines a structural prior with anomaly-aware measurements for radio-map reconstruction. Its frozen DDPM prior models macroscopic propagation from structural maps, while sparse vehicle-scattering anomalies require guided reverse inference.

  • III. PROPOSED RADIOVIL FRAMEWORK: RadioVIL combines a structural prior with anomaly-aware measurements in a constrained bilevel pipeline.This pipeline is presented in Fig. 2.
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: The framework uses structural environment maps H as conditioning information for its DDPM backbone.The backbone is denoted fθ∗(x_t, t, H).
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: The offline-trained DDPM models macroscopic radio propagation using building layouts and boundaries.These structural features are explicitly encoded in the frozen prior.
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: The frozen prior captures large-scale distance-dependent attenuation and shadowing.Its structural conditioning provides the basis for these propagation patterns.
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: The DDPM prior cannot natively represent unpredictable, highly sparse scattering anomalies caused by dynamic vehicles.This limitation distinguishes environmental structure from vehicle-induced effects.
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: Sparse measurements y are used to guide the reverse inference G(·) because the structural prior alone lacks vehicle anomalies.The passage identifies direct sparse-measurement guidance as necessary for the inference process.
  • A. Prior Adaptation and Layer-wise Deterministic Mapping: The framework accepts antenna or structural layouts, a physical observability mask, degraded sparse measurements, and complete radio maps as target outputs.These input modalities and the reconstruction target are illustrated in Fig. 1.

B. Anomaly-Aware Inversion Operator

RadioVIL introduces a sparse deviation variable to separate smooth environmental structure from high-frequency vehicle-scattering residuals in sparse radio measurements. Its layer-wise inversion uses an L1-regularized objective so the deviation explains localized anomalies, including when vehicle pixels are unobserved.

  • Anomaly-aware decomposition: The sparse deviation variable ν decouples high-frequency vehicle-scattering residuals from the smooth DDPM prior.Sparse measurements contain scattering-induced residuals that the smooth prior alone cannot represent.
  • Anomaly-aware decomposition: The model represents measurements as a smooth background fθ∗(z, t, H) plus a scattering footprint K(ν) from sparse hidden anomalies.The latent variable z adapts the smooth generative manifold to observed large-scale propagation, while ν explains residuals inconsistent with that prior.
  • Layer-wise inversion objective: The layer-wise objective minimizes measurement mismatch between observed samples and the combined smooth background-plus-anomaly prediction.The mismatch is evaluated through the measurement mask M and noise scale σ_n.
  • Sparse anomaly recovery: The L1 penalty suppresses arbitrary residual fitting and encourages localized vehicle-induced anomalies inferable from boundary samples even when vehicle pixels are unobserved.Thus, ν functions as a sparse inverse variable rather than hallucinating missing pixels.

C. Iterative Inversion and Zero-Shot Extraction

The method iteratively optimizes latent states and sparse defects within the diffusion process, then extracts vehicle locations from the resulting anomaly map. Thresholding and connected-component analysis yield physically isolated vehicle clusters without paired labels or auxiliary detection networks.

  • Iterative inversion: At each diffusion step, the procedure initializes z and ν, then jointly updates them through K gradient-descent iterations.The optimized state is passed to standard sampling to compute x_t−1, while the optimized defect is accumulated into the anomaly estimate.
  • Anomaly extraction: After the diffusion loop, the method produces a high-fidelity background radio map R_est = x_0 and an aggregated anomaly map.A threshold τ converts the anomaly map into a binary activation mask B_veh.
  • Vehicle localization: Connected Component Analysis filters isolated noise pixels, with each robust cluster C_k corresponding to a physical vehicle.Vehicle coordinates (x_k, y_k) are computed as anomaly-intensity-weighted centers of mass.
  • Zero-shot extraction: The procedure enables zero-shot vehicle localization from physically isolated anomalies without paired labels or auxiliary detection networks.Localization is performed directly from the extracted anomaly structure.

IV. EXPERIMENTAL RESULTS AND ANALYSIS · A. Experimental Setup

RadioVIL is evaluated on RadioMapSeer under leakage-controlled sparse sensing, using static environmental conditioning while withholding vehicle locations for evaluation. The study compares supervised, adversarial, and zero-shot diffusion baselines using reconstruction and vehicle-localization metrics, with specified optimization settings.

  • A. Experimental Setup: RadioMapSeer evaluation excludes vehicle locations from conditioning inputs, retaining static environmental cues and sparse observations to avoid target leakage.The conditioning map includes buildings, roads, greenbelts, base-station cues, infrastructure, coordinates, a greenbelt observation band, and 5%–10% randomly sampled free-space observations.
  • A. Experimental Setup: RadioVIL is compared with RadioUNet, RME-GAN, and RePaint, representing supervised regression, adversarial generation, and zero-shot diffusion inpainting.RadioUNet is optimized for pixel-wise Mean Squared Error (MSE), RME-GAN hallucinates missing data through adversarial training, and RePaint forcibly replaces known pixels during reverse timesteps.
  • A. Experimental Setup: RadioVIL’s outer diffusion loop follows the sampling trajectory while its inner DMILO loop minimizes J_t through L1-regularized optimization of ν.This optimization decouples the smooth background manifold f_θ*(z) from sparse scattering anomalies for target localization.
  • A. Experimental Setup: Table I reports quantitative performance for radio-map inpainting and zero-shot vehicle localization.Relative gain is computed against RePaint, and RePaint’s P90 relative gain is omitted because its P90 uses only 11.44% successful detections.
  • A. Experimental Setup: PSNR, SSIM, and LPIPS assess radio-map inpainting quality, while Recall and centroid-distance criteria assess hidden-vehicle localization.A prediction is a True Positive when its Euclidean distance from a ground-truth centroid is within 10 meters; distance metrics are computed exclusively over true positives.
  • A. Experimental Setup: Adam performs layer-wise gradient descent with learning rates of 2 × 10−2 for z and 1 × 10−3 for ν.The mediating latent state z and sparse deviation variable ν use separate learning rates.
  • A. Experimental Setup: The implementation uses K = 200 inner iterations and 5 reverse timesteps in the outer diffusion loop.These settings define the inner optimization depth and outer diffusion trajectory length.
  • A. Experimental Setup: Physical constraints are enforced with sparsity regularization fixed at λ = 0.1.The supplied implementation description also states that the anomaly activation setting follows this constraint, but the passage is truncated before its value.

B. Radio Map Inpainting Performance

RadioVIL outperforms zero-shot RePaint in radio-map inpainting while preserving vehicle-scattering signatures that conventional metrics and smoothing-prone baselines can obscure. Its layer-wise DMILO optimization separates environmental structure from scattering deviations, enabling more physically faithful reconstructions.

  • Metric limitations: 25.50 dB PSNR for RME-GAN and 25.13 dB for RadioUNet exceed RadioVIL’s 24.30 dB, but PSNR penalizes high-frequency deviations.This metric preference encourages overly smoothed reconstructions that erase vehicle-scattering footprints.
  • Comparison with RePaint: 24.30 dB vs. 23.35 dB PSNR, higher SSIM, and LPIPS 0.0587 establish RadioVIL’s advantage over RePaint.RadioVIL achieves the best reported LPIPS score of 0.0587 in this comparison.
  • Comparison with RePaint: RePaint blends sharp physical scattering boundaries into the background because its visual-semantic harmonization treats them as noise.This explains RePaint’s limited preservation of critical target signatures.
  • RadioVIL mechanism: Layer-wise DMILO optimization separates the smooth environmental prior from scattering deviations, preserving physically meaningful target signatures.RadioVIL is therefore designed to retain high-frequency vehicle-scattering information rather than harmonize it away.

C. Vehicle Localization Performance

RadioVIL enables robust zero-shot localization of hidden vehicles from sparse, strictly masked radio observations by preserving high-frequency scattering anomalies, outperforming conventional reconstruction methods.

  • Vehicle Localization Performance: RadioUNet and RME-GAN detect no hidden vehicles, each yielding an F1-Score of 0.
  • Vehicle Localization Performance: RePaint achieves 11.44% Recall and an F1-Score of 0.1765, providing only marginal vehicle detection.
  • Vehicle Localization Performance: RePaint’s lower 90th Percentile Distance of 6.41m versus 7.40m is computed over only 11.44% successful detections, so Recall and F1-Score remain essential.RadioVIL bounds localization error for a much larger portion of hidden vehicles and preserves structural anomalies that RePaint suppresses.
  • Vehicle Localization Performance: RadioVIL achieves an F1-Score of 0.6799, recalls 75.20% of hidden vehicles, and reaches an average localization error of 3.31 meters.These results use zero-shot extraction from the isolated anomaly map under sparse and strictly masked observations.
  • Vehicle Localization Performance: RadioVIL bridges visual image completion and anomaly-aware physical environment sensing.

D. Robustness to Environmental Factors · V. CONCLUSION

RadioVIL remains robust across increasingly complex environments and transmission ranges, while its frozen DDPM prior and layer-wise DMILO optimization preserve electromagnetic anomalies for zero-shot vehicle localization. The framework therefore reframes sparse radio-map reconstruction as a prior-guided physical inverse problem for future 6G ISAC sensing.

  • D. Robustness to Environmental Factors: An 8.3% decrease in vehicle detection F1-Score occurs in dense urban scenes with more than 30 buildings despite stronger blockage, shadowing, and multipath effects.Performance shows no clear monotonic degradation as environmental clutter increases.
  • D. Robustness to Environmental Factors: Radio map inpainting and vehicle localization remain relatively stable across transmission ranges from 50m to 800m.This stability indicates reliance on global environmental structural priors rather than local signal intensity alone.
  • V. CONCLUSION: RadioVIL jointly performs radio map inpainting and zero-shot hidden vehicle localization from sparse measurements using an anomaly-aware diffusion framework.The framework combines a frozen DDPM prior with layer-wise DMILO optimization.
  • V. CONCLUSION: RadioVIL reformulates radio map reconstruction as a prior-guided physical inverse problem rather than conventional pixel-level completion.Its frozen diffusion design preserves high-frequency electromagnetic signatures.
  • V. CONCLUSION: The preserved high-frequency electromagnetic signatures avoid suppression observed with regression-based, GAN-based, and semantic diffusion inpainting baselines.These signatures are central to anomaly-aware reconstruction.
  • V. CONCLUSION: CCA-based vehicle localization uses the isolated anomaly map without paired vehicle labels or auxiliary detection networks, enabling zero-shot sensing under strict masking conditions.The results support sparse wireless measurements as a sensing mechanism for future 6G ISAC systems when guided by structural generative priors and anomaly-aware inverse optimization.
Loading 2608.16167v1…