Wavelet-based Multi-View Fusion of 4D Radar Tensor and Camera for Robust 3D Object Detection

TL;DR

This work tackles the sparsity and limited semantic richness of 4D radar by fusing raw radar tensors with monocular camera data through a two-pronged approach: (i) Wavelet Attention Mixture-of-Experts (WA-MoE) enhances multi-scale radar representations by exploiting joint spatial-frequency structure, and (ii) Geometry-guided Progressive Fusion (GPF) alignment combines geometry priors with image semantics in a coarse-to-fine, two-stage fusion pipeline. The method decomposes the radar tensor into range-azimuth (RA) and elevation-azimuth (EA) views, enabling efficient yet information-rich cross-modal fusion within a multi-scale framework and iterative detection refinement. On the K-Radar dataset, WRCFormer achieves state-of-the-art 3D detection performance, particularly under adverse weather, demonstrating robust all-weather perception by leveraging raw 4D radar information and precise cross-modal alignment. This approach offers a practical, scalable path for reliable autonomous-driving perception by preserving radar’s rich cues while maintaining computational efficiency.

Abstract

4D millimeter-wave (mmWave) radar has been widely adopted in autonomous driving and robot perception due to its low cost and all-weather robustness. However, its inherent sparsity and limited semantic richness significantly constrain perception capability. Recently, fusing camera data with 4D radar has emerged as a promising cost effective solution, by exploiting the complementary strengths of the two modalities. Nevertheless, point-cloud-based radar often suffer from information loss introduced by multi-stage signal processing, while directly utilizing raw 4D radar data incurs prohibitive computational costs. To address these challenges, we propose WRCFormer, a novel 3D object detection framework that fuses raw radar cubes with camera inputs via multi-view representations of the decoupled radar cube. Specifically, we design a Wavelet Attention Module as the basic module of wavelet-based Feature Pyramid Network (FPN) to enhance the representation of sparse radar signals and image data. We further introduce a two-stage query-based, modality-agnostic fusion mechanism termed Geometry-guided Progressive Fusion to efficiently integrate multi-view features from both modalities. Extensive experiments demonstrate that WRCFormer achieves state-of-the-art performance on the K-Radar benchmarks, surpassing the best model by approximately 2.4% in all scenarios and 1.6% in the sleet scenario, highlighting its robustness under adverse weather conditions.

Figures (10)

• Figure 1: Overall comparison of state-of-the-art models, EchoFusion liu2023echoes, DPFT fent2024dpft, and ASF paek2025availability, with WRCFormer (Ours) by fusion of camera and 4D radar tensors on K-Radar dataset (v1.0 and v2.0) paek2022k.
• Figure 2: Visualization of one-frame data, including an image, elevation-azimuth map, range-azimuth map and the radar point clouds (Cartesian in BEV).
• Figure 3: The architecture of proposed WRCFormer, which has two inputs: one-frame monocular image and one-frame 4D radar tensor. WRCFormer has one output of 3D bounding box.
• Figure 4: The structure of Wavelet Attention MoE (WA-MoE).
• Figure 5: The structure of Geometry-driven Semantic Alignment, which is the stage-1 of Geometry-guided Progressive Fusion (GPF).