search

jhkohpav / MGTANet

MIT License
53 stars 5 forks source link

About sequence processing #9

Closed SteveJMao closed 11 months ago

SteveJMao commented 1 year ago

Hello，great appreciation to your great work！ I was wondering about the processing pipline about long-term bev aggregation. I noticed that you apply 3 past frames to do long term bev temporal alignment and I wondered how you deal with the first two frames which do not have previous 3 frames. Are you just skip the first two frames or apply different operations? Looking forward to hearing from you soon!

junhyung-SPALab commented 11 months ago

Hello. I'm co-author Junhyung Lee. Sorry for the late reply. As you mentioned, if you use a total of 3 frames including the target frame(denoted as t), the learning method for the first and second frame of each scene in nuScenes dataset is as follows.

  1. Multi frame composition to train the first frame: {t-2: the first frame, t-1: the first frame, t: the first frame}
  2. Multi frame composition to train the second frame: {t-2: the first frame, t-1: the first frame, t: the second frame}

Any additional questions are welcome !!

SteveJMao commented 11 months ago

Thank you for your replying！Now I understand it!