RadarGaussianDet3D: Gaussian 3D Radar Detector

• RadarGaussianDet3D is a 4D radar-based 3D detector that replaces traditional pillar encoders with a Gaussian-centric design for enhanced feature representation.
• It employs a Point Gaussian Encoder to convert sparse radar points into dense 3D Gaussian primitives using fast Gaussian splatting for robust BEV construction.
• The method introduces a Box Gaussian Loss that jointly optimizes box center, scale, and orientation, leading to improved detection accuracy and high throughput.

RadarGaussianDet3D is a single-modality 4D-radar 3D object detector that replaces classical pillar encoders and independent box regression with a Gaussian-centric design. It treats radar points as 3D Gaussian primitives and 3D bounding boxes as 3D Gaussian distributions, using fast Gaussian splatting for bird’s-eye-view (BEV) feature construction and a distribution-based box loss for training. The method is organized around a Point Gaussian Encoder (PGE), a CenterPoint-style BEV detector, and a Box Gaussian Loss (BGL), with the stated aim of improving representation quality and inference speed for sparse and noisy 4D automotive radar point clouds (Xiong et al., 19 Sep 2025).

1. Problem setting and design rationale

RadarGaussianDet3D is formulated for 4D automotive radar, where each point carries 3D spatial information after calibration together with angular information, Doppler, and intensity or radar cross-section. In the formulation adopted by the method, the practical difficulty is not the availability of BEV-based detection architectures, but the mismatch between those architectures and radar point-cloud statistics. Existing 4D radar-based 3D detectors are described as relying heavily on pillar encoders for BEV feature extraction, where each point contributes to only a single BEV grid, resulting in sparse feature maps and degraded representation quality. They are also described as optimizing bounding box attributes independently, which leads to sub-optimal detection accuracy, while inference speed may still fail to meet the real-time requirement on vehicle-mounted embedded devices (Xiong et al., 19 Sep 2025).

The proposed response is explicitly Gaussian-centered at two levels. First, radar points are converted into Gaussian primitives whose learned scale and rotation allow each point to influence multiple BEV cells through Gaussian splatting rather than a single pillar assignment. Second, predicted and ground-truth boxes are converted into 3D Gaussian distributions, so localization, scale, and orientation are optimized jointly through a distributional distance rather than by independent scalar penalties. This design preserves the familiar CenterPoint-style detection pipeline while replacing its input representation and part of its regression objective (Xiong et al., 19 Sep 2025).

A useful clarification is that RadarGaussianDet3D is not a radar densification method in the sense of generating LiDAR-like point clouds, nor is it a scene-level Gaussian field method that first builds a sparse set of object-centric latent Gaussians. Its basic unit is the observed radar point, which is expanded into a Gaussian primitive and then rasterized into BEV. This makes the method structurally closer to a fast radar encoder than to diffusion-based enhancement or full neural scene reconstruction.

2. Point Gaussian Encoder

The Point Gaussian Encoder consists of local feature aggregation, global feature aggregation, Gaussian attribute prediction, and BEV Gaussian splatting. The input is a radar point cloud

$\{(p_i, f_i)\}_{i=1}^N,$

where $p_i \in \mathbb{R}^3$ are 3D coordinates and $f_i \in \mathbb{R}^{C_{\text{raw}}}$ are raw point features. PGE first enriches these features locally through a PointNet-like neighborhood operator. For point $i$, with spherical neighborhood $\mathcal{N}_i$ of radius $r$, the local feature is

$$f^i_{\textup{LFA}}= \frac{1}{|\mathcal N_i|} \sum_{j\in\mathcal N_i} \mathtt{Linear}\big(\mathtt{Concat}([f_j,p_j-p_i])\big).$$

This explicitly injects relative offsets $p_j-p_i$, following the logic of point-based local geometry encoding (Xiong et al., 19 Sep 2025).

Global interactions are then modeled by point-level self-attention. With raw features $f$, the global feature aggregation stage is

$$\begin{aligned} f_1 &= \mathtt{Linear}(f),\ Q,K,V &= \mathtt{MLP}(\mathtt{LayerNorm}(f_1)),\ f_2 &= \mathtt{SelfAttn}(Q,K,V)+f_1,\ f_{\textup{GFA}} &= \mathtt{FFN}(\mathtt{LayerNorm}(f_2))+f_2. \end{aligned}$$

Because radar points are sparse and many pillars would contain only one point, this point-level attention is used to capture global structure without relying on dense pillar occupancy (Xiong et al., 19 Sep 2025).

From the concatenated raw, local, and global features, PGE predicts Gaussian attributes: $p_i \in \mathbb{R}^3$0 where $p_i \in \mathbb{R}^3$1 is the scale vector, $p_i \in \mathbb{R}^3$2 is a rotation quaternion, and $p_i \in \mathbb{R}^3$3 is a view-independent feature vector. The Gaussian primitive for each radar point is

$p_i \in \mathbb{R}^3$4

with $p_i \in \mathbb{R}^3$5 and $p_i \in \mathbb{R}^3$6. The covariance is constructed from scale and rotation as

$p_i \in \mathbb{R}^3$7

The design choice to fix $p_i \in \mathbb{R}^3$8 and $p_i \in \mathbb{R}^3$9 is important: later ablations show that predicting position offsets and opacity harms performance in sparse radar settings (Xiong et al., 19 Sep 2025).

For BEV rasterization, the 3D Gaussian is projected by parallel projection: $f_i \in \mathbb{R}^{C_{\text{raw}}}$0 so that

$f_i \in \mathbb{R}^{C_{\text{raw}}}$1

The BEV feature map $f_i \in \mathbb{R}^{C_{\text{raw}}}$2 is then rendered by modified 3D Gaussian splatting: $f_i \in \mathbb{R}^{C_{\text{raw}}}$3 Instead of assigning each point to one grid cell, each point contributes to all overlapping BEV cells according to its projected Gaussian density. The resulting feature map is therefore denser and smoother than a pillar scatter map (Xiong et al., 19 Sep 2025).

3. Box Gaussian Loss and detection formulation

RadarGaussianDet3D keeps a CenterPoint-style BEV backbone and CenterHead detection head, but replaces purely attribute-wise box regression with Box Gaussian Loss. A 3D box

$f_i \in \mathbb{R}^{C_{\text{raw}}}$4

is converted to a Gaussian by setting

$f_i \in \mathbb{R}^{C_{\text{raw}}}$5

$f_i \in \mathbb{R}^{C_{\text{raw}}}$6

$f_i \in \mathbb{R}^{C_{\text{raw}}}$7

The scaling factor $f_i \in \mathbb{R}^{C_{\text{raw}}}$8 is class-dependent: $f_i \in \mathbb{R}^{C_{\text{raw}}}$9 for pedestrians and cyclists, and $i$0 for cars and trucks (Xiong et al., 19 Sep 2025).

Given a predicted Gaussian $i$1 and a ground-truth Gaussian $i$2, the box loss is the average Kullback–Leibler divergence

$i$3

with

$i$4

This formulation couples center error, scale mismatch, and orientation mismatch in a single distance. The first term is a Mahalanobis distance weighted by the ground-truth box geometry, so a fixed translational offset is penalized differently for large and small boxes. The trace and determinant terms penalize covariance mismatch, which amounts to mismatch in box extent and orientation (Xiong et al., 19 Sep 2025).

BGL is added to the original CenterHead regression objective: $i$5 with $i$6 in the reported experiments. Classification remains unchanged. At inference time, BGL introduces no additional cost, because it is only used during training (Xiong et al., 19 Sep 2025).

4. Empirical results, ablations, and efficiency

RadarGaussianDet3D is evaluated on TJ4DRadSet and View-of-Delft. On TJ4DRadSet, it reports 3D AP values of 26.69 for car, 28.18 for pedestrian, 65.84 for cyclist, and 19.63 for truck, for a 3D mAP of 35.08%; the corresponding BEV mAP is 41.98%. On View-of-Delft validation, as a radar-only detector, it reports EAA AP values of 40.7 for car, 42.4 for pedestrian, and 73.0 for cyclist, for an EAA mAP of 52.0%, and ROI AP values of 71.2, 51.7, and 89.0, for an ROI mAP of 70.6% (Xiong et al., 19 Sep 2025).

Dataset	Accuracy summary	Runtime
TJ4DRadSet	35.08% 3D mAP, 41.98% BEV mAP	43.5 Hz on Tesla V100
View-of-Delft val	52.0% EAA mAP, 70.6% ROI mAP	83.2 Hz on Tesla V100

These results place the method close to or at the top of radar-only baselines while preserving high throughput. On TJ4DRadSet, MAFF-Net records a slightly higher 3D mAP of 35.38%, but RadarGaussianDet3D is reported at 43.5 Hz on Tesla V100, whereas MAFF-Net is reported at 17.9 Hz on RTX 4090. On View-of-Delft, RadarGaussianDet3D is second best among radar-only methods and is described as very close to LiDAR-assisted SCKD while remaining radar only (Xiong et al., 19 Sep 2025).

The ablation study isolates the contribution of each Gaussian component. Starting from a CenterPoint-Pillar baseline at 29.04% 3D mAP on TJ4DRadSet, adding Gaussian splatting alone yields 30.89%, adding GS+LFA yields 32.20%, adding GS+GFA yields 32.33%, combining GS+LFA+GFA as full PGE yields 33.50%, and adding BGL produces the full 35.08%. This indicates that the gain is not attributable to a single modification; the BEV scattering mechanism, point-feature aggregation, and Gaussian box supervision each contribute distinct improvements (Xiong et al., 19 Sep 2025).

A second ablation addresses Gaussian parameterization. Predicting both position offsets and opacity yields 32.39% 3D mAP and 39.72% BEV mAP; predicting opacity only yields 33.30% and 40.14%; predicting neither, and thus fixing $i$7 and $i$8, yields the best result at 33.50% and 41.15%. This is one of the method’s more notable findings: in sparse radar, additional Gaussian degrees of freedom can degrade accuracy rather than increase it (Xiong et al., 19 Sep 2025).

Efficiency is also analyzed at the module level. For local feature aggregation on TJ4DRadSet, a traversal implementation requires 177.9 ms and 202.6 MB, broadcasting and masking require 3.9 ms and 3981.6 MB, and the proposed indexing-and-scattering implementation requires 0.5 ms and 202.6 MB. This supports the article’s central characterization of PGE as a low-latency replacement for pillar encoding rather than a heavy auxiliary module (Xiong et al., 19 Sep 2025).

5. Subsequent extensions and research lineage

Later work explicitly positions RadarGaussianDet3D as the origin of the Gaussian Splatting-based Point Gaussian Encoder. RCGDet3D describes RadarGaussianDet3D as a 4D radar-only 3D object detector that introduced PGE, and states that its radar branch is based on RadarGaussianDet3D for its effectiveness and low latency. In that formulation, the radar branch is composed of a Point Gaussian Encoder and the BEV backbone neck, and the detector head is CenterPoint-style with the same Box Gaussian Loss added for bounding-box regression (Xiong et al., 20 May 2026).

RCGDet3D proposes two changes to the original encoder. First, Ray-centric PGE predicts Gaussian attributes in ray-aligned coordinate systems before unifying them to BEV space, rather than predicting them directly in ego coordinates. Second, a Semantic Injection module uses deformable attention to inject image features into radar point features before Gaussian prediction. In the radar-only setting, this produces what the paper calls RadarGaussianDet3D+. On View-of-Delft, RadarGaussianDet3D is reported at 52.0% EAA mAP, 70.6% ROI mAP, and 83.2 Hz on V100, while RadarGaussianDet3D+ reaches 53.6%, 70.8%, and 77.5 Hz. On TJ4DRadSet, the same replacement improves 3D mAP from 35.08% to 36.71% and BEV mAP from 41.98% to 43.07%, at essentially unchanged speed (Xiong et al., 20 May 2026).

A separate branch of research uses Gaussian primitives more globally. RaGS, for example, is a radar-camera detector that uses a cascaded pipeline to initialize, refine, and rasterize a limited number of 3D Gaussians as object-centric anchors, rather than expanding each observed radar point into a Gaussian primitive for BEV encoding (Bai et al., 26 Jul 2025). This suggests a broader distinction inside Gaussian-based radar detection: one line uses per-point Gaussian BEV encoding, as in RadarGaussianDet3D, while another uses scene-level or object-level Gaussian fields as the primary latent representation.

6. Limitations, adjacent methods, and outlook

RadarGaussianDet3D does not claim to solve the fundamental limitations of radar sensing. Its own discussion identifies radar sparsity and noise as remaining fundamental issues, and its ablations show that over-parameterized Gaussian primitives can hurt rather than help. The method’s future directions are explicitly multi-modal fusion and temporal modeling, which indicates that the presented formulation is best understood as a strong radar-only baseline and encoder rather than a complete account of Gaussian radar perception (Xiong et al., 19 Sep 2025).

Adjacent work points to at least two complementary directions. One direction is input-level radar enhancement before detection. R$i$9D models the residual between radar BEV and LiDAR BEV and applies sigma-adaptive regional guidance, with the explicit interpretation that it can serve as a front-end for a RadarGaussianDet3D pipeline by turning raw mmWave radar measurements into dense, LiDAR-like 3D structures while preserving radar-specific uncertainty (Li et al., 10 Jan 2026). HyperDet, in a different setting, constructs a task-aware hyper 4D radar point cloud by spatio-temporal aggregation, cross-sensor consensus validation, and foreground-focused diffusion, and then feeds that refined point cloud into standard LiDAR-oriented detectors without architectural modifications (Xiao et al., 12 Feb 2026). Both approaches target the same bottleneck that RadarGaussianDet3D addresses at the encoder level: sparse, unstable radar geometry.

A second direction is broader Gaussian radar scene modeling. Gaussian representations have already been used for radar-camera 3D object detection beyond the radar-only setting (Bai et al., 26 Jul 2025), for 4D radar-camera SLAM and dynamic masking in outdoor environments (Xiao et al., 20 Nov 2025), for 4D radar-inertial odometry through freely positioned 3D Gaussians and multi-hypothesis scan matching (Amodeo et al., 2024), and for radar Gaussian splatting as a scene-reconstruction and data-synthesis engine (Kung et al., 2 Jun 2025). This broader literature does not reduce RadarGaussianDet3D to a special case, but it does place it inside a larger shift toward explicit Gaussian primitives as intermediate representations for radar geometry, uncertainty, and downstream reasoning.

In that context, RadarGaussianDet3D occupies a specific and influential position: it demonstrates that a Gaussian formulation can be inserted directly into a high-throughput radar-only detector, improving both BEV feature density and box optimization without abandoning the operational simplicity of CenterPoint-style detection. Subsequent work has largely treated that combination—fast per-point Gaussian encoding plus distribution-aware supervision—as a reusable foundation rather than a terminal design (Xiong et al., 19 Sep 2025).