Revisiting Radar Perception With Spectral Point Clouds

Hamza Alsharif; Jing Gu; Pavol Jancura; Satish Ravindran; Gijs Dubbelman

cs.CVApr 9, 2026

Revisiting Radar Perception With Spectral Point Clouds

Authors: Hamza Alsharif, Jing Gu, Pavol Jancura, Satish Ravindran, Gijs Dubbelman

Organizations: Eindhoven University of Technology · NXP Semiconductors

Abstract

Radar perception models are trained with different inputs, from range-Doppler spectra to sparse point clouds. Dense spectra are assumed to outperform sparse point clouds, yet they can vary considerably across sensors and configurations, which hinders transfer. In this paper, we provide alternatives for incorporating spectral information into radar point clouds and show that, point clouds need not underperform compared to spectra. We introduce the spectral point cloud paradigm, where point clouds are treated as sparse, compressed representations of the radar spectra, and argue that, when enriched with spectral information, they serve as strong candidates for a unified input representation that is more robust against sensor-specific differences. We develop an experimental framework that compares spectral point cloud (PC) models at varying densities against a dense range-Doppler (RD) benchmark, and report the density levels where the PC configurations meet the performance of the RD benchmark. Furthermore, we experiment with two basic spectral enrichment approaches, that inject additional target-relevant information into the point clouds. Contrary to the common belief that the dense RD approach is superior, we show that point clouds can do just as well, and can surpass the RD benchmark when enrichment is applied. Spectral point clouds can therefore serve as strong candidates for unified radar perception, paving the way for future radar foundation models.

Figures & tables

Figure 1: Comparing radar perception approaches on RADIal [ 21 ] object detection. Point cloud models can compete with dense RD spectra models, and spectral enrichment pushes their performance further.

Figure 2: RD spectrum visualization with corresponding ground truth bounding boxes. The RD frame is incoherently integrated across Rx and visualized in log scale. The sensor uses simultaneous-Tx phase coding that interleaves per-Tx signatures along the Doppler axis (DDMA), yielding repeating patterns.

Figure 3: RD and PC training pipelines . The radar network can be trained with either RD spectra frames or spectral point clouds. In the RD pipeline, an implicit learnable RD to RA transformation is applied. In the PC pipeline, structure is imposed by a projection of latent features onto a BEV grid.

Figure 4: Spectral point cloud data equivalence to RD . A spectral point cloud is equivalent in data to a sparse RD frame with cells from peak set $\mathcal{S}(\tau)$ , after CFAR-based data reduction.

Figure 5: PC F1 score trend, with axis ablations. The angle amplitude

A_{\Theta}

axis yields a

2.0

point average improvement over spatial coordinates

[r,a,e]

, and the Doppler bin

d

axis yields a

3.6

point average improvement. Including both axes results in a

4.6

point average improvement. At

30.8\%

density, full-axes point clouds surpass the RD benchmark.

Modality	Density (%)	mIoU (%)
RD benchmark	100	79.4
PC	0.5	78.3
PC	1.6	79.6
PC	5.7	80.6
PC	19.0	80.7
PC	30.8	81.0

Table 1: Segmentation results for various densities.

Figure 6: PC enrichment scheme performance. RD neighborhood enrichment reaches the RD benchmark at

17.2\%

density but can slightly reduce F1 at very low densities, whereas angle spectrum enrichment reaches the benchmark at

19.0\%

density without low-density degradation. SpectralPillars reaches the benchmark earliest, at

7.2\%

density.

M.	Ds. (%)	F1	SP. Ds. (%)	SP. F1	FPS
FF	100	91.5	-	-	169
PP	0.7	81.0	0.7	74.3	279
PP	1.6	86	1.8	86.0	276
PP	5.7	90.2	5.6	90.9	264
PP	8.8	90.2	7.2	91.6	258
PP	19.0	90.5	17.2	91.3	232

Table 2: Model throughput for various densities. FF: FFTRadNet, PP: PointPillars, SP: SpectralPillars. M.: Model, Ds.: Density.

Model	Mode	Params ( $10^{6}$ )	AP ( $\%$ )	AR ( $\%$ )	F1 ( $\%$ )	RE ( $m$ )	AE (°)
FFTRadNet [ 21 ]	RD	3.8	96.8	82.2	88.9	0.11	0.17
T-FFTRadNet [ 7 ]	RD	9.6	89.5	89.6	89.5	0.15	0.12
SparseRadNet [ 28 ]	RD	6.9	96.0	91.8	93.8	0.13	0.10
Pixor [ 21 , 33 ]	PC	6.9	96.5	32.3	48.4	0.17	0.25
PointPillars [ 10 ]	PC	2.4	88.7	93.5	91.0	0.13	0.21
SpectralPillars (Ours)	PC	2.4	94.3	91.5	92.9	0.12	0.22

Table 3: Comparison with state-of-the-art. Bold text denotes the best result, and underlined denotes the second best. The last three rows are results that we have produced based on the default density (

5.7\%

), whereas the rest are taken from their respective papers. Pixor, considerably underperforms compared to RD results.

Figure 7: SpectralPillars compared to sparse RD. Average F1 gain over sparse RD is +1.3 points.

Millimeter-Wave Radar

Millimeter-wave radar uses reflected high-frequency radio waves to sense motion, distance, and scene structure in settings where cameras may fail.

See all 51 papers

Depth-Semantic Alignment and Affinity-Guided Fusion for Structured Radar Point Cloud Generation
Jun 25, 2026Amjad Hussain, Xin Qiu, Fuyuan Ai +4Camera-Radar Fusion Point Cloud Completion
Zhejiang University, China
SD4R: Sparse-to-Dense Learning for 3D Object Detection with 4D Radar
Feb 24, 2026Xiaokai Bai, Jiahao Cheng, Songkai Wang +53D Object Detection Point Cloud Learning
College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, China · School of Automotive Studies, Tongji University, Shanghai 201804, China · Ningbo Global Innovation Center, Zhejiang University, Ningbo 315100, China
IRGNN: Efficient Invariant Radar Graph Neural Network for Radar Point Cloud Object Detection
Aug 14, 2026Xiao Guo, Wanke Xia, Lili Yang +13D Object Detection Autonomous Driving Perception
China Agricultural University · Tsinghua University

Jun 25, 2026J/K move · Enter open · S save

Depth-Semantic Alignment and Affinity-Guided Fusion for Structured Radar Point Cloud Generation

Amjad Hussain, Xin Qiu, Fuyuan Ai, Yuchen Tan+3

Zhejiang University, China

Camera-Radar Fusion Point Cloud Completion Point Cloud Generation

Open paper AskPDF arXiv

Point clouds are an important carrier of three-dimensional spatial information, and their quality directly affects the performance of downstream perception tasks such as object detection and tracking. However, millimeter-wave radar point clouds are typically sparse, noisy, and structurally incomplete. To address these limitations, this paper proposes a multimodal point cloud generation method based on vision-radar fusion. The proposed method leverages image semantic information to impose structural constraints and achieve spatial alignment for radar point clouds, while incorporating a sparse completion strategy to enhance point density and recover missing structures. The generated point clouds are further evaluated in object detection and tracking tasks. Experimental results demonstrate that the proposed method effectively improves point cloud quality and enhances the detection accuracy and robustness of perception models in complex environments, providing a practical solution for multisensor point cloud generation and intelligent perception systems.

Abstract

Millimeter-Wave Radar

Explore similar work

Depth-Semantic Alignment and Affinity-Guided Fusion for Structured Radar Point Cloud Generation

SD4R: Sparse-to-Dense Learning for 3D Object Detection with 4D Radar

IRGNN: Efficient Invariant Radar Graph Neural Network for Radar Point Cloud Object Detection