How can I ensure the GFE module is differentiable?

[ChiefGodMan]

After reading this paper, the core idea of one is how to transform bev seg mask to vectorized points feature. But I don’t understand how to ensure the GFE module, aka Thresholding-Max Survival-FPS, differentiable and efficiency ? And after we extract vectorized points from instance mask, why do we need to encode them to decode the result? In my opinion we can keep the extracted points as pred result.

[SehwanChoi0307]

The module is designed to explicitly extract the geometric information of each instance at the point level and pass it to the PQG Query. Moreover, since the results extracted by the GFE do not have sequential information, it is not possible to know the direction of the lane. Therefore, the final decoding process is performed.