search

wangyx-chn / BIDNet

TGRS paper: Extracting Building Footprints in SAR Images via Distilling Boundary Information From Optical Images
0 stars 0 forks source link

Questions about the learning BIDNet #1

Open cch-meiple opened 3 months ago

cch-meiple commented 3 months ago

Hi, thank you for your work! I am attempting to train the BIDNet and would greatly appreciate your assistance with the following questions:

  1. The SN6 SAR imagery consists of 4-channel, 32-bit data. How did you convert this to 8-bit?
  2. Am I correct in understanding that to train the teacher model, I should use train.py with RGB_res18_train_teacher.yaml, and for training the student model, I should use train_dist.py with SAR_res18_train_distill.yaml?
  3. What exactly is the dist_cfg variable in SAR_res18_train_distill.yaml, and how can I generate the data for it?

Thank you for your answer.

wangyx-chn commented 1 month ago
  1. you can take a normalization operator for conversion
  2. yes
  3. it refers to the distillation layer and the path of the pre-results generated by the RGB network. You can first infer the RGN images with the teacher model. Then train the student model with the RGB edge map and so on.