search

hustvl / MapTR

[ICLR'23 Spotlight & IJCV'24] MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction
MIT License
1.09k stars 168 forks source link

code details #24

Open koughua opened 1 year ago

koughua commented 1 year ago

Hi，thanks for your great work, i'm interested in it.

I noticed that you predict a bbox in the code, and use the loss function to constrain it during the training process. What is the role of the predicted output bbox? I am very confused, and it seems that you did not mention it in the article.

Looking forward to your reply

LegendBC commented 1 year ago

Thanks for your interest! We don't use it during training. Adding the bbox loss to further constrain the polyline points range does not help the map learning, so we set the bbox loss weight to 0 as seen in our config files.

yihongXU commented 1 year ago

Hi, It seems also that you do not consider Box IOU during matching in training? Does it mean that instance level matching is in practice point-level matching. I mean, they are performed jointly in the code and only pts_cost and cls_cost are non zero.