RadSplat: 3D Gaussian Splatting & Radiance

• RadSplat is a framework that leverages explicit 3D Gaussian primitives and radiance fields to synthesize scenes with high fidelity and efficiency.
• It integrates radiance field pretraining with differentiable splatting, rigorous pruning, and sensor-specific optimizations for real-time performance.
• Applications include photorealistic rendering, radar-based SLAM, and wireless radiomap extrapolation, achieving notable improvements in speed and quality.

RadSplat refers to a family of frameworks that leverage 3D Gaussian Splatting in combination with radiance fields, radar or radio signal modeling, or physics-informed priors to achieve high-fidelity, efficient, and robust scene synthesis or environmental modeling. While the specific meaning of "RadSplat" varies by context and community, it encompasses advances in photorealistic rendering, sensor simulation, wireless radiomap extrapolation, and radar data synthesis, unified by an explicit 3D Gaussian point-based representation and optimized mapping between physical measurements and 3D scene structure.

1. Overview and Definition

RadSplat originally denotes a real-time, radiance field-informed Gaussian Splatting approach for robust view synthesis. In this setting, RadSplat fuses the high-fidelity and robustness of volumetric radiance fields (such as NeRFs) with the inference and rendering scalability of 3D Gaussian Splatting (3DGS). This framework has been adapted in parallel for radar and radio-frequency-centric applications, including radar-camera SLAM, radar sensor data synthesis, and wireless radiomap extrapolation.

Across these variants, RadSplat employs a collection of explicit 3D Gaussian primitives characterized by learned spatial parameters (mean, covariance, opacity), often augmented by modality-specific attributes (such as color SH coefficients, reflectivity, noise estimates) (Niemeyer et al., 2024, Kung et al., 2 Jun 2025, Wang et al., 18 Feb 2025, Xiao et al., 20 Nov 2025). The systems commonly share the following principles:

• Scene geometry and appearance are represented explicitly as a finite set of anisotropic 3D Gaussians.
• Rendering and data association are implemented via differentiable splatting-based compositing, enabling direct optimization of Gaussian parameters through gradient-based learning.
• Domain adaptation is achieved by fusing sensor-specific priors (radiance fields, radar signal models, wireless propagation) into the optimization or initialization process.
• Scalability, interpretability, and data efficiency are enhanced through point pruning, hierarchical management, and explicit modeling of sensor or environmental noise.

2. Radiance Field-Informed Gaussian Splatting for Real-Time Rendering

The canonical RadSplat system addresses the limitations of both radiance-field-based and point-based real-time rendering (Niemeyer et al., 2024). Standard radiance fields perform volume rendering of a function $f_\theta(x, d) \rightarrow (\sigma, c)$, which, while photorealistic, is computationally prohibitive for real-time applications due to stochastic volumetric integration. In contrast, 3D Gaussian Splatting enables GPU-accelerated rasterization but suffers from brittle point optimization and unbounded model growth.

RadSplat fuses these paradigms by:

1. Pretraining a Radiance Field Prior: Utilizing a method such as Zip-NeRF (with per-image GLO latents) to establish a data-robust volumetric representation, trained via photometric losses over batches of rays.
2. Initialization and Supervision of 3DGS: Initializing millions of Gaussians along median-depth sample rays informed by the pretrained radiance field, inheriting color, position, and spatial scale, and then supervising their optimization against "cleaned" NeRF renderings with explicit SSIM and $L_2$ losses.
3. Pruning and Visibility Filtering: Systematically pruning Gaussians with negligible integrated ray contributions, and implementing post hoc viewpoint-based masks via k-means clustering of camera poses and cluster-level importance scoring.

This hybridization achieves robust photorealistic synthesis at interactive rates, yielding 900+ FPS on commodity GPUs for large complex scenes, with a substantial reduction (up to 10×) in the number of active Gaussians compared to conventional splatting (Niemeyer et al., 2024).

3. Radar and Radiomap Variants of Gaussian Splatting

In radar-centric and wireless communication domains, RadSplat principles underpin several systems that extend the explicit Gaussian splatting paradigm to non-optical modalities:

• Radar-Vision 3DGS SLAM (Rad-GS): This system fuses 4D radar (range, azimuth, elevation, Doppler) and camera data to generate a globally consistent, differentiable Gaussian map suitable for outdoor SLAM (Xiao et al., 20 Nov 2025). Dynamic masking is achieved via Doppler-based indexing and semantic segmentation (EfficientSAM), and both front-end (keyframe-synchronized) and back-end (unsynchronized) frames jointly optimize Gaussian parameters and camera poses.
• Radar Gaussian Splatting for Data Synthesis (RadarSplat): Applied to autonomous driving, RadarSplat models radar phenomenology—including receiver saturation, multipath, and speckle—by extending each 3D Gaussian with explicit radar reflectivity, noise probability, and view-dependent response, and optimizes these against observed polar radar images. Splat-based rendering directly synthesizes raw radar outputs with physically-consistent artifacts, outperforming prior NeRF-style radar models in PSNR, SSIM, and geometric accuracy (Kung et al., 2 Jun 2025).
• RadSplatter for Wireless Radiomap Extrapolation: 3DGS is repurposed to model the multipath environments upon which radio-frequency signals propagate. Each scatterer is encoded as a Gaussian, with learned attenuation and spherical harmonics coefficients. Camera-free splatting is achieved through learned 3D-to-beamspace mappings via MLPs, and domain-specific recursive fine-tuning exploits sparse ground-truth measurements for robust extrapolation (Wang et al., 18 Feb 2025).

4. Optimization, Regularization, and Scalability

Optimization in RadSplat and its derivatives leverages differentiable compositing: for each pixel or sensor ray, Gaussians are sorted and composited front-to-back via the accumulation of opacities and parameterized outputs (color, radar intensity, or RSS). Backpropagation through all model components enables robust convergence and modular incorporation of multiple loss terms, e.g., photometric error, SSIM, occupancy map constraints, roughness priors, and explicit regularizations on Gaussian scale or selection matrices (Niemeyer et al., 2024, Wang et al., 18 Feb 2025, Xiao et al., 20 Nov 2025).

To maintain scalability and prevent model bloat:

• Ray-Contribution-Based Pruning: Importance scores (e.g., $h(p_i) = \max_{views, r} \alpha_i^r \tau_i^r$) are used to cull low-impact Gaussians, yielding up to 10× model compression without quality loss (Niemeyer et al., 2024).
• Visibility Filtering: Precomputed cluster-wise visibility lists drastically reduce computation at test time and allow for efficient, scalable rendering of large scenes (Niemeyer et al., 2024).
• Hierarchical Octree Management: For SLAM and large-scale mapping, aggressive octree-based merging and splitting of Gaussian primitives control memory and promote detailed resolution where warranted by local geometry or data density (Xiao et al., 20 Nov 2025).

5. Quantitative Performance and Experimental Validation

Across variants, RadSplat establishes new baselines in both rendering speed, perceptual quality, and scene reconstruction accuracy.

Method	SSIM	PSNR	LPIPS	FPS	# Gaussians (M)
Ours Light (RadSplat, MipNeRF360)	0.826	27.56	0.213	907	0.37
3DGS	0.815	27.20	0.214	251	3.16
Zip-NeRF	0.836	28.54	0.177	0.25	—
SMERF	0.818	27.99	0.211	228	—

In radar-centric evaluation (RadarSplat) (Kung et al., 2 Jun 2025):

• PSNR: +3.4 dB over Radar Fields (26.06 vs. 22.66)
• SSIM: ×2.6 improvement (0.51 vs. 0.20)
• RMSE: –40% in 3D geometry (1.81m vs. 3.03m)
• Geometric inference saturates at $2 \times 10^4$ Gaussians for typical automotive-scale scenes.

For RF radiomap extrapolation (RadSplatter) (Wang et al., 18 Feb 2025):

• MAE (dB) on synthetic (Shanghai): 7.564 vs. 9.867 (NeRF$^2$ baseline)
• Inference $\sim 4\times$ faster than strong neural and interpolation baselines.

SLAM evaluation (Rad-GS, (Xiao et al., 20 Nov 2025)):

• PSNR and localization error on kilometer-scale scenes outperforming previous camera/LiDAR 3DGS SLAM.

6. Limitations and Directions for Future Work

While RadSplat and its extensions deliver state-of-the-art performance across domains, several limitations are noted:

• Elevated training costs compared to single-representation baselines due to radiance field pretraining and combined optimization steps ($\sim$2 h end-to-end for the main pipeline).
• Residual PSNR/LPIPS gaps in extreme, in-the-wild or district-scale captures, indicating room for further scalability.
• In radar and RF contexts, most current models are static-scene only; dynamic object integration and wave-penetration occlusion remain open research challenges.
• For radiomap extrapolation, modeling large-scale spatial discontinuities and topology-sensitive propagation requires further advances in Gaussian selection and parameterization.

Planned research directions include training and inference acceleration, scene-graph extensions for dynamic environments, regularization for wave-penetration effects, as well as domain-specific simulator-to-real transfer pipelines for robust perception system development.

7. Context within the Broader Literature

RadSplat represents a unifying framework in the evolution of neural scene representations that balances explicit geometric parameterization (Gaussians), physically-informed signal modeling, and computational efficiency. Its direct antecedents and related approaches include classic neural radiance fields (NeRFs), Zip-NeRF, 3D Gaussian Splatting, and domain-adapted neural mapping frameworks for sensor simulation and SLAM. By incorporating sensor-specific priors and explicit data-driven pruning/optimization strategies, RadSplat provides a flexible template adaptable to photorealistic rendering, radar simulation, wireless coverage modeling, SLAM, and beyond (Niemeyer et al., 2024, Kung et al., 2 Jun 2025, Wang et al., 18 Feb 2025, Xiao et al., 20 Nov 2025).