arxiv: 2604.25945 · v1 · submitted 2026-04-17 · 📡 eess.SP · cs.AI

Recognition: unknown

Planar Gaussian Splatting with Bilinear Spatial Transformer for Wireless Radiance Field Reconstruction

Jinghan Zhang , Xitao Gong , Qi Wang , Richard A. Stirling-Gallacher , Giuseppe Caire

Authors on Pith no claims yet

Pith reviewed 2026-05-10 08:01 UTC · model grok-4.3

classification 📡 eess.SP cs.AI

keywords wireless radiance fieldGaussian splattingspatial power spectrumbilinear spatial transformerelectromagnetic couplingradio frequency reconstructionspatial spectrum synthesis

0 comments

The pith

Planar Gaussian splatting rendered on angular spectra with bilinear attention reconstructs wireless radiance fields more accurately than NeRF or prior GS approaches.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops BiSplat-WRF to learn continuous representations of radio-frequency behavior across 3D space and direction from measurements. It replaces standard 3D Gaussian splatting with 2D planar Gaussians placed at 3D coordinates but rendered directly onto the angular domain of spatial power spectra. A bilinear spatial transformer then links these primitives across an angular grid through attention, so that long-range electromagnetic coupling and scattering are modeled explicitly. The resulting fields are tested on the task of predicting spatial power spectra at a receiver for given transmitter locations. If the method works, it supplies a queryable model that can replace slower or less accurate neural representations in wireless system design and sensing.

Core claim

BiSplat-WRF is a planar Gaussian splatting framework in which each primitive is a 2D planar Gaussian with 3D coordinates that is rendered directly onto the angular domain of the SPS. The bilinear spatial transformer aggregates relations among these primitives on an angular grid and applies attention to capture global EM coupling and mutual scattering, thereby enforcing physically consistent long-range interactions. On the spatial spectrum synthesis task the method records higher SSIM than both NeRF-based and earlier GS-based baselines; ablation experiments confirm that the BST component accounts for a measurable fraction of the gain. A scaled-up BiSplat-WRF+ variant further raises SSIM at a

What carries the argument

BiSplat-WRF framework consisting of 2D planar Gaussians rendered directly onto the SPS angular domain together with a bilinear spatial transformer that performs attention-based aggregation of inter-primitive electromagnetic relations.

If this is right

Higher SSIM scores on spatial spectrum synthesis compared with NeRF and earlier Gaussian-splatting baselines.
Explicit modeling of global EM coupling and mutual scattering through the bilinear spatial transformer attention mechanism.
A larger BiSplat-WRF+ variant that trades higher compute cost for still larger SSIM gains.
Removal of unnecessary 3D-to-2D projection steps while preserving the expressiveness of Gaussian primitives.
Validation through ablation studies that isolate the contribution of the bilinear spatial transformer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same planar-plus-attention pattern could be tested on acoustic or optical wave fields that obey similar propagation physics.
Integration with real-time channel sounding hardware would allow closed-loop validation of predicted spectra against live measurements.
The angular-grid attention structure might reduce the number of required measurement locations compared with purely local splatting methods.
If the BST attention can be made sparse, the approach could scale to larger environments without quadratic cost growth.

Load-bearing premise

Rendering 2D planar Gaussians straight onto the angular domain and letting bilinear attention link them is enough to reproduce the global electromagnetic coupling and scattering physics of real wireless channels.

What would settle it

A side-by-side comparison on measured data from a rich-scattering indoor environment where BiSplat-WRF predictions deviate systematically from ground-truth spatial power spectra in directions known to contain strong multipath components not representable by the planar primitives.

Figures

Figures reproduced from arXiv: 2604.25945 by Giuseppe Caire, Jinghan Zhang, Qi Wang, Richard A. Stirling-Gallacher, Xitao Gong.

**Figure 1.** Figure 1: Illustration of spatial power spectrum where ym,n represents the complex signal received at antenna element (m, n), and ∆σm,n(ϕ, θ) is the geometric phase shift relative to the reference antenna (0, 0). B. 3D Gaussian Splatting 3D Gaussian Splatting models a scene using a collection of 3D anisotropic Gaussian primitives distributed throughout the space. This explicit representation offers superior comput… view at source ↗

**Figure 2.** Figure 2: Overall framework of BiSplat-WRF. The architecture integrates static and dynamic signal modeling with a BST for global EM coupling. Signal values [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Illustration of BST. It takes the concatenation of GS-PE and TX-PE as input and projects to joint encoded local feature. A bilinear feature aggregation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Illustration of bilinear feature aggregation in BST, where the region [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Cumulative Distribution Function (CDF) of SSIM for various SPS [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: The median SSIM values of the methods with different training data [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

read the original abstract

Wireless radiance field (WRF) reconstruction aims to learn a continuous, queryable representation of radio frequency characteristics over 3D space and direction, from which specific quantities, such as the spatial power spectrum (SPS) at a receiver given a transmitter position, can be predicted. While Gaussian splatting (GS)-based method has surpassed Neural Radiance Fields (NeRF)-based method for this task, existing adaptations largely transplant vision pipelines, limiting physical interpretability and accuracy. We introduce BiSplat-WRF, a planar GS framework that retains the expressiveness of 3D GS while removing unnecessary projections and incorporating global EM coupling and mutual scattering among primitives. Each primitive is a 2D planar Gaussian with 3D coordinates, rendered directly on the angular domain of the SPS. A bilinear spatial transformer (BST) aggregates inter-primitive relations on an angular grid and, via attention, captures long-range electromagnetic dependencies, thereby enforcing globally aware EM interactions that reflect the complex physics of the wireless environment. On spatial spectrum synthesis task, BiSplat-WRF surpasses NeRF-based and prior GS-based baselines with respect to the Structural Similarity Index (SSIM); comprehensive ablation studies validate the contribution of BST. We also provide a larger BiSplat-WRF+ variant that further increases SSIM at a higher computation cost, serving as a strong reference for future studies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BiSplat-WRF adapts planar Gaussian splatting and a bilinear transformer to wireless radiance fields with reported SSIM gains over NeRF and prior GS baselines, but the physical grounding of the BST remains thin.

read the letter

The paper introduces BiSplat-WRF, which renders 2D planar Gaussians with 3D coordinates directly onto the angular domain of the spatial power spectrum and adds a bilinear spatial transformer to aggregate inter-primitive relations via attention. This setup is positioned as capturing long-range EM dependencies and mutual scattering better than transplanted vision pipelines. The main empirical result is higher SSIM on the spatial spectrum synthesis task, backed by ablations that attribute gains to the BST component and a heavier BiSplat-WRF+ variant offered as a reference point. These elements are genuinely new relative to the cited NeRF and GS work in the wireless setting. The direct angular rendering avoids some unnecessary 3D projection steps and fits the SPS output format cleanly. The ablations provide a straightforward check on the transformer addition. The central soft spot is the interpretation that the attention mechanism enforces globally aware EM interactions reflecting complex wireless physics. Nothing in the architecture explicitly enforces wave-equation consistency, diffraction, multipath interference, or frequency-dependent behavior. The observed SSIM lift could arise from added capacity and optimization rather than physical fidelity, which matches the stress-test concern. Without checks against known propagation cases or Maxwell-consistent simulations, the physics claim stays aspirational. This work is for researchers in wireless signal processing and sensing who want continuous, queryable RF representations. It shows coherent adaptation of existing tools and supplies concrete numbers plus ablations, so it merits a serious referee even with the noted gap on physical grounding. Recommendation: send it to peer review.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces BiSplat-WRF, a planar Gaussian splatting framework for wireless radiance field (WRF) reconstruction. Each primitive is a 2D planar Gaussian with 3D coordinates rendered directly onto the angular domain of the spatial power spectrum (SPS). A bilinear spatial transformer (BST) aggregates inter-primitive relations on an angular grid via attention to capture long-range electromagnetic dependencies. The method claims to surpass NeRF-based and prior GS-based baselines in SSIM on the spatial spectrum synthesis task, with ablation studies validating the BST contribution; a larger BiSplat-WRF+ variant is also provided as a reference.

Significance. If the SSIM gains are attributable to the BST's modeling of global EM coupling rather than added capacity alone, the work could advance physically interpretable representations for wireless channel modeling and spectrum prediction, bridging vision-based splatting with electromagnetic principles and providing a stronger baseline for future WRF studies.

major comments (2)

[§3] §3 (Method): The central claim that BST 'enforces globally aware EM interactions that reflect the complex physics' and captures 'global EM coupling and mutual scattering' is load-bearing for interpreting the SSIM improvements as physically grounded. No derivation is provided showing how the bilinear attention on the angular grid enforces consistency with Maxwell's equations, diffraction, multipath interference, or frequency-dependent behavior; without this, gains may reduce to empirical fitting.
[§4] §4 (Experiments): The abstract states that 'comprehensive ablation studies validate the contribution of BST' and that BiSplat-WRF surpasses baselines in SSIM, but no quantitative results, error bars, dataset details, or statistical tests are referenced. This undermines assessment of whether the reported superiority is robust or sensitive to post-hoc choices in architecture or training.

minor comments (1)

[Abstract] Abstract: The title and abstract introduce 'BiSplat-WRF' and 'BST' without immediate expansion; adding a parenthetical definition on first use would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below with proposed revisions to improve the manuscript's clarity and rigor.

read point-by-point responses

Referee: [§3] §3 (Method): The central claim that BST 'enforces globally aware EM interactions that reflect the complex physics' and captures 'global EM coupling and mutual scattering' is load-bearing for interpreting the SSIM improvements as physically grounded. No derivation is provided showing how the bilinear attention on the angular grid enforces consistency with Maxwell's equations, diffraction, multipath interference, or frequency-dependent behavior; without this, gains may reduce to empirical fitting.

Authors: We agree that the manuscript provides no formal derivation from Maxwell's equations or explicit consistency with diffraction, multipath, or frequency dependence. The BST is a data-driven attention mechanism on the angular grid, motivated by the directional and long-range dependencies inherent in electromagnetic wave propagation. The SSIM gains are empirical, as shown by the ablations. We will revise §3 to moderate the language, framing the BST as capturing global interactions in a manner inspired by EM principles rather than enforcing physical laws, and add a short discussion of these motivations along with a limitations note on the empirical nature of the improvements. revision: yes
Referee: [§4] §4 (Experiments): The abstract states that 'comprehensive ablation studies validate the contribution of BST' and that BiSplat-WRF surpasses baselines in SSIM, but no quantitative results, error bars, dataset details, or statistical tests are referenced. This undermines assessment of whether the reported superiority is robust or sensitive to post-hoc choices in architecture or training.

Authors: The full manuscript reports SSIM comparisons, ablation results, and dataset specifications in §4 and the tables. However, the abstract omits specific numbers, and error bars plus statistical tests are absent. We will revise the abstract to include key quantitative SSIM gains, expand dataset details, and add error bars from multiple runs plus statistical significance tests (e.g., t-tests) in the experimental section to better demonstrate robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: modeling choices validated empirically against external baselines

full rationale

The paper introduces a new planar Gaussian splatting architecture with BST attention for wireless radiance field reconstruction. Claims that the components 'enforce globally aware EM interactions' and 'reflect the complex physics' are interpretive descriptions of the model design rather than a derivation chain that reduces to fitted inputs or self-citations by construction. Performance gains are demonstrated via direct comparisons to NeRF and prior GS baselines plus ablation studies on BST, which are independent external benchmarks. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The framework is self-contained as an empirical ML proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into parameters or axioms; the approach assumes Gaussian primitives and attention can model EM physics but introduces BST as a new mechanism without external validation.

axioms (1)

domain assumption Planar 2D Gaussians with 3D coordinates can retain expressiveness of 3D GS while enabling direct angular-domain rendering for SPS
Invoked to justify the core representation change from vision pipelines

invented entities (1)

Bilinear Spatial Transformer (BST) no independent evidence
purpose: Aggregates inter-primitive relations on angular grid and captures long-range electromagnetic dependencies via attention
New component introduced to enforce globally aware EM interactions

pith-pipeline@v0.9.0 · 5560 in / 1377 out tokens · 49426 ms · 2026-05-10T08:01:40.201555+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 6 canonical work pages · 1 internal anchor

[1]

A survey of various propagation models for mobile communication,

T. K. Sarkar, Z. Ji, K. Kim, A. Medouri, and M. Salazar-Palma, “A survey of various propagation models for mobile communication,”IEEE Antennas and propagation Magazine, vol. 45, no. 3, pp. 51–82, 2003

2003
[2]

The design and applications of high-performance ray-tracing simulation platform for 5G and beyond wireless communications: A tutorial,

D. He, B. Ai, K. Guan, L. Wang, Z. Zhong, and T. K ¨urner, “The design and applications of high-performance ray-tracing simulation platform for 5G and beyond wireless communications: A tutorial,”IEEE communications surveys & tutorials, vol. 21, no. 1, pp. 10–27, 2018

2018
[3]

Sionna: An Open-Source Library for Next-Generation Physical Layer Research,

J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An open-source library for next-generation physical layer research,”arXiv preprint arXiv:2203.11854, 2022

work page arXiv 2022
[4]

Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,

S. Jiang, Q. Qu, X. Pan, A. Agrawal, R. Newcombe, and A. Alkhateeb, “Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,”IEEE Open Journal of the Communications Society, 2025

2025
[5]

Generalizable neural ray tracing towards physics-informed intelligent channel modeling,

K. Bian, M. Tao, and S. Sun, “Generalizable neural ray tracing towards physics-informed intelligent channel modeling,” in2025 IEEE/CIC International Conference on Communications in China (ICCC). IEEE, 2025, pp. 1–6

2025
[6]

Neural reflectance fields for radio-frequency ray tracing,

H. Jia, X. Chen, Y . Wei, Y . Sun, and Y . Pi, “Neural reflectance fields for radio-frequency ray tracing,” inGLOBECOM 2024-2024 IEEE Global Communications Conference. IEEE, 2024, pp. 4226–4231

2024
[7]

NeRF2: Neural radio-frequency radiance fields,

X. Zhao, Z. An, Q. Pan, and L. Yang, “NeRF2: Neural radio-frequency radiance fields,” inProceedings of the 29th Annual International Con- ference on Mobile Computing and Networking, 2023, pp. 1–15

2023
[8]

3D Gaussian splatting for real-time radiance field rendering

B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3D Gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

2023
[9]

A Survey on 3D Gaussian Splatting

G. Chen and W. Wang, “A Survey on 3D Gaussian Splatting,” 2025. [Online]. Available: https://arxiv.org/abs/2401.03890

work page internal anchor Pith review Pith/arXiv arXiv 2025
[10]

3D Gaussian Splatting as a New Era: A Survey,

B. Fei, J. Xu, R. Zhang, Q. Zhou, W. Yang, and Y . He, “3D Gaussian Splatting as a New Era: A Survey,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 8, pp. 4429–4449, 2025

2025
[11]

WRF-GS: Wireless radiation field reconstruction with 3D Gaussian splatting,

C. Wen, J. Tong, Y . Hu, Z. Lin, and J. Zhang, “WRF-GS: Wireless radiation field reconstruction with 3D Gaussian splatting,” inIEEE IN- FOCOM 2025-IEEE Conference on Computer Communications. IEEE, 2025, pp. 1–10

2025
[12]

Scalable 3D Gaussian splatting- based RF signal spatial propagation modeling,

K. Yang, W. Du, and M. Srivastava, “Scalable 3D Gaussian splatting- based RF signal spatial propagation modeling,” inProceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems, 2025, pp. 680–681

2025
[13]

Neural representation for wireless radiation field reconstruction: A 3D Gaussian splatting approach

C. Wen, J. Tong, Y . Hu, Z. Lin, and J. Zhang, “Neural representation for wireless radiation field reconstruction: A 3D Gaussian splatting approach,”arXiv preprint arXiv:2412.04832, 2024

work page arXiv 2024
[14]

Rasterizing Wireless Radiance Field via Deformable 2D Gaussian Splatting,

M. Liu, C. Zhang, Q. Yang, Y . Cao, Y . Xu, Y . Xu, S. Sun, M. Dai, and Y . Guan, “Rasterizing Wireless Radiance Field via Deformable 2D Gaussian Splatting,”arXiv preprint arXiv:2506.12787, 2025

work page arXiv 2025
[15]

RF-PGS: Fully-Structured Spatial Wireless Channel Representation with Planar Gaussian Splatting,

L. Zhang, Z. Li, and H. Sun, “RF-PGS: Fully-structured Spatial Wireless Channel Representation with Planar Gaussian Splatting,”arXiv preprint arXiv:2508.16849, 2025

work page arXiv 2025
[16]

Survey on 6G frontiers: Trends, applications, re- quirements, technologies and future research,

C. De Alwis, A. Kalla, Q.-V . Pham, P. Kumar, K. Dev, W.-J. Hwang, and M. Liyanage, “Survey on 6G frontiers: Trends, applications, re- quirements, technologies and future research,”IEEE Open Journal of the Communications Society, vol. 2, pp. 836–886, 2021

2021
[17]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

2017
[18]

V oxelRF: V oxelized radiance field for fast wireless channel modeling,

Z. Zeng, S. Sun, M. Tao, Y . Xu, and X. Yu, “V oxelRF: V oxelized radiance field for fast wireless channel modeling,”arXiv preprint arXiv:2507.09987, 2025

work page arXiv 2025