SDEF-BEV: spatial-aware dual-expert radar-camera fusion for robust BEV 3D object detection

Engineering notes

Experiments on the nuScenes dataset demonstrate that SDEF-BEV achieves an NDS of 57.1% and an mAP of 45.7%, showing competitive performance.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Bird's-eye view (BEV) radar-camera fusion is an economical and robust solution for 3D perception in autonomous driving. However, existing methods struggle with spatial misalignment between modalities and lack the ability to dynamically adjust the contribution of each modality across different spatial regions in complex scenarios. To address these issues, we propose SDEF-BEV, a novel spatial-aware dual-expert fusion network. At its core, SDEF-BEV adopts a parallel dual-path fusion architecture. The Spatial-Aware Dual-Expert Fusion (SDEF) module adaptively learns optimal feature fusion strategies via two specialized experts (semantic-enhanced and geometry-enhanced) and a spatial gating network, generating location-wise fusion weights. Meanwhile, a parallel Channel and Spatial Fusion (CSF) path preserves rich spatial context through convolutional inductive biases. The outputs of these two paths are aggregated to form unified and information-rich fused features. Experiments on the nuScenes dataset demonstrate that SDEF-BEV achieves an NDS of 57.1% and an mAP of 45.7%, showing competitive performance. Extensive ablation studies further validate the effectiveness of the parallel SDEF architecture, and the overall approach exhibits strong robustness, particularly under challenging adverse weather conditions.

Links and sources

• PDF from original source

Comments