کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

UASP: Uncertainty-Aware Soft PointPainting for Robust LiDAR-Camera Fusion in PointPillars

نویسندگان :

Morteza Badakhshan¹ Mohammad Azim Karami²

1- دانشگاه علم و صنعت ایران 2- دانشگاه علم و صنعت ایران

کلمات کلیدی :

LiDAR-camera fusion،3D object detection،PointPillars،PointPainting،uncertainty aware،entropy gating،misalignment robustness

چکیده :

We propose UASP, a PointPillars-based fusion framework that treats image semantics as probabilistic cues rather than hard labels. UASP introduces two mechanisms: (1) Soft PointPainting, which aggregates segmentation predictions within a K×K image neighborhood to reduce sensitivity to projection residuals; and (2) Entropy-Gated Fusion, which down-weights semantic features when predictive uncertainty is high. To keep inference practical, We precompute per-pixel semantic probabilities offline and the PointPillars detector remains unchanged at runtime. On KITTI validation, UASP improves over the LiDAR-only PointPillars baseline by +3.12 BEV mAP and +3.11 3D mAP (Moderate). On the KITTI test benchmark, UASP improves over PointPillars by +2.84 BEV mAP and +2.80 3D mAP (Moderate), with the largest gains on Pedestrian and Cyclist across difficulty levels. These results indicate that uncertainty-aware soft painting increases robustness to projection and semantic noise while preserving the efficiency of pillar-based detection.