UA-CF: Uncertainty-Aware Camera-LiDAR Fusion for Robust 3D Object Detection Under Adverse Weather

Engineering notes

Because large-scale public datasets could not be downloaded in the local artifact environment, we provide a transparent synthetic BEV-proposal benchmark rather than unverifiable leaderboard claims.

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

Robust 3D object detection remains a critical bottleneck for autonomous driving under fog, rain, and snow because camera and LiDAR streams do not fail in the same way. Cameras provide dense semantic cues but lose contrast under scattering and low visibility, whereas LiDAR preserves metric structure but suffers from sparsification, backscatter, and spurious returns under precipitation and dense aerosols. This paper proposes UA-CF, an uncertainty-aware camera-LiDAR fusion framework that dynamically allocates feature-level weights according to modality reliability. The method combines camera semantic features, LiDAR bird's-eye-view geometric features, label-free sensor-quality estimates, inverse-uncertainty fusion, and a calibration-aware objective. Unlike fixed fusion, UA-CF is designed to avoid over-trusting a degraded modality while retaining complementary information from both streams. Building on the positive gated vision-LiDAR motivation, our work focuses on a camera-LiDAR setting and introduces explicit reliability weighting for adverse-weather 3D detection. Because large-scale public datasets could not be downloaded in the local artifact environment, we provide a transparent synthetic BEV-proposal benchmark rather than unverifiable leaderboard claims. Across five random seeds, UA-CF obtains 0.949 AP overall, compared with 0.916 for equal fusion and 0.905 for logistic concatenation. Under fog, UA-CF reaches 0.988 AP and assigns 0.842 mean weight to LiDAR; under rain, it reaches 0.959 AP and shifts 0.683 mean weight to camera features. These reproducible results support the central claim that uncertainty-aware fusion improves robustness under asymmetric sensor degradation while preserving clear-weather performance.

Links and sources

Comments