fundamentalvision/BEVFormer

Employs spatiotemporal transformer architecture with spatial cross-attention across multi-camera views and temporal self-attention for recurrent BEV fusion, enabling unified representations for both 3D detection and map segmentation tasks. Built on MMDetection3D framework with multiple configuration variants (tiny to base) trading memory for performance, from 6.5GB to 28.5GB, and includes BEVFormerV2 with improved backbone integration reaching 56.9% NDS on nuScenes—9 points above prior camera-only methods.

4,356 stars. No commits in the last 6 months.