search

justchenhao / STANet

official implementation of the spatial-temporal attention neural network (STANet) for remote sensing image change detection
BSD 2-Clause "Simplified" License
423 stars 110 forks source link

Inaccurate Results with PAM Pre-Trained Weights in demo.py #109

Open Keerthana-chinta opened 2 months ago

Keerthana-chinta commented 2 months ago

Hello,

I am experiencing issues with the accuracy of the results when using the PAM pre-trained weights while running demo.py. Despite following the instructions and loading the PAM pre-trained weights (pam_net_F.pth and pam_net_A.pth), the output does not seem to reflect the expected performance. Could you please verify whether the pre-trained weights are compatible with the current implementation in demo.py? Are there any known issues or additional steps required to get accurate results with the PAM pre-trained weights? Any guidance on improving the accuracy or resolving this issue would be greatly appreciated. Thank you for your support and for providing this valuable resource to the community!

ravmarcinMT commented 5 days ago

I tested the pam_net_F.pth and pam_net_A.pth in PAM pre-trained folder, and find out that the scores suggest that they are not pre-trained (or are very early checkpoint). I got the recall 5.6%. I assume that you must retrain the model using instructions from README