References & Citations
Computer Science > Computer Vision and Pattern Recognition
Title: BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection
(Submitted on 21 Jun 2022 (v1), last revised 30 Nov 2022 (this version, v2))
Abstract: In this research, we propose a new 3D object detector with a trustworthy depth estimation, dubbed BEVDepth, for camera-based Bird's-Eye-View (BEV) 3D object detection. Our work is based on a key observation -- depth estimation in recent approaches is surprisingly inadequate given the fact that depth is essential to camera 3D detection. Our BEVDepth resolves this by leveraging explicit depth supervision. A camera-awareness depth estimation module is also introduced to facilitate the depth predicting capability. Besides, we design a novel Depth Refinement Module to counter the side effects carried by imprecise feature unprojection. Aided by customized Efficient Voxel Pooling and multi-frame mechanism, BEVDepth achieves the new state-of-the-art 60.9% NDS on the challenging nuScenes test set while maintaining high efficiency. For the first time, the NDS score of a camera model reaches 60%.
Submission history
From: Zheng Ge [view email][v1] Tue, 21 Jun 2022 03:21:18 GMT (4457kb,D)
[v2] Wed, 30 Nov 2022 11:28:11 GMT (2913kb,D)
Link back to: arXiv, form interface, contact.