search

YaoChengTang / Sparse-Point-Guided-3D-Lane-Detection

Sparse Point Guided 3D Lane Detection
MIT License
5 stars 1 forks source link

There might be something different between codes and paper #2

Open betterCZ101 opened 10 months ago

betterCZ101 commented 10 months ago

In the paper, it's mentioned that "Perspective Transformer is used in coarse-level lane detection to extract the dense BEV features and detect coarse 3D lanes at the lowest resolution." But in the codes, the "projs = self.pers_tr(input, frontview_features, _M_inv)" , and the projs have the four different levels of resolution of BEV features, so the training parameters does not decrease a lot. Maybe I misunderstood, Could you have some explanation?

YaoChengTang commented 10 months ago

We use the variable 'use_all_fea' to control the generation of BEV features in the code of PerspectiveTransformer, as shown in Ours and PersFormer.

betterCZ101 commented 10 months ago

thank for quick reply！