Loading...
Training-Free Late Fusion across Geometry and BEV for Edge-Deployable LiDAR-Camera 3D Perception
Zhang, Yixuan
Zhang, Yixuan
Abstract
Autonomous driving depends on accurate and reliable 3D perception, and LiDAR–camera fusionis central to that goal. However, most multisensor fusion pipelines limit portability and hinderreal-world deployment. To understand the strengths of dominant LiDAR–camera fusionparadigms compared to single-sensor perception, we present a unified benchmark comparingthree representative 3D perception pipelines: CenterPoint (LiDAR only), a geometry-level fusionmodel (MVP) that injects image cues into point space, and a bird’s-eye view (BEV)-level fusionmodel (BEVFusion) that aggregates multimodal features in BEV. We further propose a training-free late fusion module that applies consensus and de-noising across the strongest models’predictions to improve detection quality across operating conditions. Our results show that (i)geometry-level and BEV-level fusion offer complementary strengths rather than a single winner:BEVFusion achieves the highest overall detection accuracy, while MVP provides more precisespatial localization; (ii) a late fusion stage can combine the advantages of both paradigms and issuitable for real-time deployment on in-vehicle edge hardware, without requiring retraining.
Description
Date
2025-12-08
Journal Title
Journal ISSN
Volume Title
Publisher
Collections
Download Dataset
Rights Holder
The author
Usage License
http://creativecommons.org/licenses/by-nc-sa/4.0/
Embargo
Research Projects
Organizational Units
Journal Issue
Keywords
Citation
Advisor
Department
Data Science