A Block-Cross-View Attention and Dynamic Iterative Depth Update for Multi-View 3D Object Detection

Detecting 3D objects through multi-view cameras is a crucial and prevalent part of autonomous driving. The improvement of 3D object detection is largely reflected in the boost of depth estimation accuracy, and how to leverage the geometry correlation among surrounding views in time and space becomes promising for accurate depth estimation. In this paper, we introduce a new multi-view 3D detector with enhanced depth estimation. Specifically, we firstly design a block-cross-view attention module with camera awareness to strengthen the ability for depth feature generation, and then we propose a dynamic iterative algorithm within the depth optimization process to reduce unrelated context noises to improve the BEV representations. Experiments on public data set nuScenes are conducted to verify the effectiveness and robustness of the proposed method.