BEVFusion: An open source autonomous driving BEV project

Multi-sensor fusion is crucial for accurate and reliable autonomous driving systems. State-of-the-art approaches are based on point-level fusion: augmenting LiDAR point clouds with camera features .
However, the camera-to-LiDAR projection discards the semantic density of camera features, hindering the effectiveness of such approaches, especially for semantically oriented tasks such as 3D scene segmentation. In this paper, we propose BEVFusion, an efficient and general multi-task multi-sensor fusion framework. This approach unifies multimodal features in a shared bird's-eye view (BEV) representation space, effectively preserving both geometric and semantic information. To achieve this, we diagnose and eliminate key efficiency bottlenecks in view translation through optimized BEV pooling, reducing latency by over 40x. BEVFusion is fundamentally task-agnostic and seamlessly supports diverse 3D perception tasks with minimal architectural changes. It establishes a new state-of-the-art on the nuScenes benchmark, achieving 1.3% mAP and NDS for 3D object detection and 13.6% mIoU for BEV map segmentation, all while reducing computational cost by 1.9x.