MR-BEV: A Multi-Resolution BEV Detection Framework with a Better Efficiency-Accuracy Trade-off

In multi-view 3D object detection, achieving good detection performance for small and distant objects in the focus view, like the front, is crucial for self-driving safety. A natural and cost-saving solution is to enhance the image resolution only for the focus view. However, this approach may lead to performance degradation due to the distribution discrepancy caused by multi-resolution inputs. This paper introduces MR-BEV (Bird's-Eye View), a framework for camera BEV-based 3D detection that incorporates a Multi-Resolution Normalization (MRN) layer to regularize the distribution of image features and mitigate the effects of multi-resolution inputs. Additionally, the framework includes a Multi-Resolution Refine Module (MRRM) and a Multi-Resolution Consistency (MRC) loss to further align the image features from different resolutions. Extensive experiments conducted on the nuScenes dataset demonstrate the effectiveness of our proposed methods. The overall NDS (NuScenes Detection Score) is improved by 12.3% (0.342 vs. 0.219) compared to a multi-resolution baseline method. In particular, the front-view NDS is raised by 19.7% for small objects and 22.6% for objects located 30 meters a way. Furthermore, MR-BEV runs 2.03 times faster (53.77 vs. 26.50 FPS, TensorRT-fp32) than the standard pipeline with high-resolution inputs for all views.