.png){kind=icon}
.png){kind=icon}
Return
Diff4RadBEV: Conditional Diffusion for 4D Radar BEV Enhancement
.png){kind=icon}
PRE
AI
DOI：10.1109/lra.2026.3726340.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Bird’s-Eye-View (BEV) representations of 4D radar scans are inherently sparse and noisy, limiting their practicality in downstream perception tasks. To address this issue, we cast 4D radar BEV enhancement as a conditional generative problem and propose Diff4RadBEV, a two-stage cascaded diffusion framework that maps raw radar BEV heightmaps to dense and fine-grained BEV representations. Specifically, Diff4RadBEV follows a recurrent diffusion-and-refinement paradigm: the BEV-conditioned Diffusion Network (BDN) first denoises and reconstructs the global scene layout, while the Refinement Network (RN) suppresses residual local artifacts and sharpens foreground structures. This coarse-to-fine design improves both layout consistency and local geometric clarity of the generation process, tailored to the sparse and noisy characteristics of radar BEV. Experiments demonstrate that Diff4RadBEV consistently improves raw radar BEV across BEV enhancement and downstream tasks, including place recognition and occupancy forecasting, while outperforming existing radar enhancement baselines. With only 4.6 M parameters and 5 sampling steps, Diff4RadBEV achieves efficient enhancement suitable for real-time deployment.
Keywords:
Representation learning
deep learning methods
4D mmWave radar
diffusion model
bird’s-eye-view
Journal
I
IF:
5.3
Papers:
1.7K
Citations:
3.9W