search

ggsDing / SCanNet

42 stars 2 forks source link

Request for train.py without L_psd #13

Closed Pangda1114 closed 7 months ago

Pangda1114 commented 7 months ago 
 Thank you for sharing the codes about train_psd.py! As you have mentioned in the paper, without using L_psd (only use the annotation of changed area), your model also performs well, so I want to try your model on my dataset which just have annotation of changed area.
 I  would appreciate it if you could provide the codes of the train.py without L_psd. Looking forward to your early reply!
ggsDing commented 7 months ago

Hi. Thank you for following our work.

Please find line 35 in 'train_psd.py'. Set 'psd_train' to False to disable pseudo labels.

Pangda1114 commented 7 months ago

Thanks to your quick response! I have mentioned 'psd_train' setting, but I find the dataloader still needs labelA and labelB. However, I just have labelB-A (changed area), how can I adapt my input to your model? It really confused me.

ggsDing commented 7 months ago

Thanks to your quick response! I have mentioned 'psd_train' setting, but I find the dataloader still needs labelA and labelB. However, I just have labelB-A (changed area), how can I adapt my input to your model? It really confused me.

I think you have two options. First, you can do multi-class CD instead of semantic CD (in this paper). Second, you can map your labels in the same form as in this paper. Note that, although multi-CD has the same goal as semantic CD, the evaluation metrics should be different. It would be unfair to use the metrics in this paper to directly evaluate the results of MCD.