Skip to content
Book Open access

ElevC2F: Elevation-Aware Coarse-to-Fine Fusion of 4D Radar and Camera for 3D Object Detection

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 40 references

Abstract

While the fusion of 4D millimeter-wave radar and camera holds significant potential for 3D object detection in autonomous driving, the extreme sparsity of radar points and the deficiency of depth information in images lead to severe cross-modal spatial misalignment. Existing two-stage methods might suffer from the irreversible loss of height information during view transformation, and the deficiency of semantic information within proposals during the refinement stage. To address these challenges, this paper proposes an elevation-aware fusion detection framework, designed to maximize the synergistic effects between 4D radar's spatial perception and the camera's semantic richness. First, at the scene-level proposal generation stage, we propose a global fusion module integrated with multi-granularity height feature optimization. This module leverages Deformable Cross-Attention (DCA) to associate radar and image semantics across the global scene and effectively mitigates the height collapse issue during 3D-to-BEV transformation by aggregating elevation features at different granularities, thereby achieving high-precision initial proposal regression. Second, at the region-level feature refinement stage, we design a refinement module featuring selective semantic feature injection. By progressively mapping fine-grained image semantics into 3D bounding box grids, this module compensates for the semantic deficiency within proposals while preserving the spatial geometric structures of 4D radar, enabling precise secondary refinement of object boundaries. Experimental results on VoD and TJ4DRadSet benchmarks demonstrate that ElevC2F achieves state-of-the-art performance, reaching 62.29% and 41.50% 3D mAP respectively, and outperforming existing methods by significant margins of up to +1.41% and +3.53%. Our code and supplementary materials will be available at https://github.com/fancj2017/ElevC2F.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.