无人驾驶ECCV 20252025
LSS: Lift, Splat, Shoot for Multi-View Perception
Jonathon Luiten, Gabriel Ghosh, Alexander Khrulev, Yang Li, Deva RamananUber ATG
摘要
LSS lifts 2D image features to 3D bird's-eye-view through explicit depth estimation for multi-view autonomous driving perception.
lift-splat-shootdepth estimation2D-to-3Dmulti-view perceptionBEV
技术细节
模型骨架
LSS + Depth Distribution + BEV Projection
编码器
Camera Backbone (EfficientNet) + Depth Encoder
解码器
BEV Splat Decoder + Detection Head (4 layers)