search

yuantianyuan01 / StreamMapNet

GNU General Public License v3.0
202 stars 17 forks source link

Performance gap between using the ego coordinate system and the lidar coordinate system. #26

Open Bio-shine opened 6 months ago

Bio-shine commented 6 months ago

Thanks for your work. I have tried streammapnet using the lidar coordinate system, and met a performance degradation about 10% in mAP. Maybe there exists bugs in my re-implementation. Have you ever tried modeling in the lidar coordinate system, why do you choose the ego-system?

And i found a bit difference existing in the generation of lane gts from the way in MapTR. Can you tell me the difference more specifically , as i do not find the details in the paper.

Thank you.

yuantianyuan01 commented 5 months ago

1) For the first question "Have you ever tried modeling in the lidar coordinate system?": No, but I suggest you re-checking the coordinate system. 2) For the second question, our implementation on NuScenes dataset should be almost the same as MapTR. We only add additional post-processing on Argoverse2 dataset.