无人驾驶ICLR 20252025
BEVFormer++: Enhanced Bird-Eye-View for Autonomous Driving
Zhiqi Li, Wenhai Wang, Hongyang Li, Jifeng DaiShanghai AI Laboratory, Nanjing University
摘要
We present BEVFormer++, an enhanced bird-eye-view representation for autonomous driving perception. BEVFormer++ improves upon the original with better temporal modeling, multi-scale features, and cross-modal fusion, achieving state-of-the-art results on 3D detection and map segmentation benchmarks.
bird-eye-viewtemporal modelingmulti-scale feature3D detectionmap segmentation
技术细节
模型骨架
BEVFormer++ Enhanced Transformer
编码器
Spatial/Temporal Cross-Attention Encoder + ResNet-50
解码器
Multi-Task BEV Decoder (detection+map+segmentation, 8 layers)