Improving Bird's Eye View based 3D object detection via Learnable Perspective View

Open Webpage

Ruibin Wang, Bowei Xing, Xianghua Ying

Published: 01 Jan 2025, Last Modified: 07 Nov 2025Expert Syst. Appl. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•A Learnable Perspective View that harnesses the pinhole camera model to learn supplementary views for Bird’s Eye View (BEV).•A novel view learning pipeline which first initializes the view parameters to approximate BEV and then learns offsets to adjust the initializations.•A differentiable sampling strategy to generate feature maps of the learned views.•Improving six existing BEV detectors under nuScenes and KITTI datasets in both LiDAR and camera modalities.•Outperforming other supplementary views with better detection and faster speed.