BEVStereo: Enhancing Depth Estimation in Multi-View 3D Object Detection with Temporal Stereo

Restricted by the ability of depth perception, all Multi-view 3D object detection methods fall into bottleneck accuracy. By constructing temporal stereo, estimation is quite reliable in indoor scenarios. However, there are two difficulties directly integrating stereo outdoor multi-view detectors: 1) The construction stereos for views results high computing costs. 2) Unable to adapt challenging In this study, we propose an effective method creating dynamically determining center and range stereo. most confident found using EM algorithm. Numerous experiments on nuScenes have shown BEVStereo's deal with complex scenarios that other stereo-based unable handle. For first time, a approach shows superiority like static ego vehicle moving objects. BEVStereo achieves new state-of-the-art camera-only track dataset while maintaining memory efficiency. Codes been released.