EVENT DOSSIER
The researchers proposed a lightweight module called Post Fusion Stabilizer (PFS), aimed at enhancing the stability of bird’s-eye view features in multi-modal 3D detection. This module operates on the intermediate BEV representation of existing detectors, improving robustness through feature statistics under stable domain offsets, suppressing regions affected by sensor degradation, and adaptive residual correction. In the nuScenes benchmark, PFS achieved advanced results in multiple failure scenarios: a 1.2% improvement in robustness for camera dropout scenarios and a 4.4% increase in mAP in low-light conditions, while consuming only 3.3 M parameters.
研究人员提出一种名为 Post Fusion Stabilizer (PFS) 的轻量级模块，用于稳定多模态 3D 检测中的鸟瞰图特征。该模块作用于现有检测器的中间 BEV 表示，通过稳定域偏移下的特征统计、抑制受传感器退化影响的区域以及自适应残差校正来增强鲁棒性。在 nuScenes 基准测试中，PFS 实现了多项失败模式下的最先进结果，其中相机脱落鲁棒性提升 +1.2%，低光性能提升 +4.4% mAP，同时仅占用 3.3 M 参数。