GenAD: How to achieve 3D position encoding of instances in autonomous driving?

The GenAD project uses a set of proxy markers to represent the 3D position of each instance around it. By using a deformable cross-attention mechanism, the project obtains updated proxy markers from the bird's-eye view feature markers. The advantage of this method is that it can not only effectively encode the 3D position of the instance, but also capture the relationship between the instances and their interaction with the surrounding environment, enhancing the system's perception and decision-making capabilities.