Late Fusion consists of a 2-stream encoder in which the RGB image and the LiDAR BEV inputs are processed independently of each other. These features are then combined via element-wise summation and passed to waypoint prediction network. This is equivalent to removing the transformer modules from TransFuser.
