SimPB: A Single Model for 2D and 3D Object Detection from Multiple Cameras by Yingqi Tang et al. explores a new paradigm in autonomous driving vision systems—a unified model that performs both 2D and 3D detection tasks from multiple camera inputs.

Main Contributions:

• SimPB uses a hybrid decoder to deal separately with 2D perspective view and 3D BEV space detection.
• An innovative approach of Dynamic Query Allocation and Adaptive Query Aggregation enables cyclic refinement of 2D and 3D interactions.
• The Query-group Attention mechanism strengthens interactions among 2D queries within each camera group.

The results showcased on the nuScenes dataset illustrate the potential of SimPB in enhancing object detection capabilities for autonomous vehicles, promising more accurate environmental perception. Authors make their code available at GitHub, encouraging community engagement and further improvement.