The gap between change detection in OSCD and the results in the paper.

danfenghong / IEEE_TPAMI_SpectralGPT

Hong, D., Zhang, B., Li, X., Li, Y., Li, C., Yao, J., Yokoya, N., Li, H., Ghamisi, P., Jia, X., Plaza, A. and Gamba, P., Benediktsson, J., Chanussot, J. (2024). SpectralGPT: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. DOI:10.1109/TPAMI.2024.3362475.

191 stars 20 forks source link

The gap between change detection in OSCD and the results in the paper. #14

Open xxxxyliu opened 7 months ago

xxxxyliu commented 7 months ago

I achieved an F1 score of only around 30 in change detection using the same dataset, while the result reported in the paper is 50. Has anyone encountered a similar issue? I would greatly appreciate any advice you could provide.

moonboy12138 commented 7 months ago

I'm not sure about your issue. Please try using several models to ensure that the data processing procedure is functioning correctly.

xxxxyliu commented 6 months ago

I'm not sure about your issue. Please try using several models to ensure that the data processing procedure is functioning correctly.

Thank you for your reply. In fact, I haven't made any changes to the data processing part. I will try to find the reason for the issue again. Thank you once more.

sebbelese commented 2 months ago

Did you figure out the configuration to use? I tend to stabilize around F1~0.47 using the repo model and settings.

csparc commented 2 months ago

I have the same issue. Stabilises around F1~0.34, using the spectralGPT+ model downloaded from here, lr=0.001. Did anyone figure this out?