z-zheng / xview2_4th_solution Goto Github PK

Hi Zhuo - I am working to replicate your results on the xBD dataset using your method. I noticed you were also releasing similar code to this solution at your ChangeOS repository. Do you have any suggestions for replicating the results posted to the xView2 evaluation leaderboard? I am able to achieve somewhat reasonable results, but localization and major/minor damage F1 performance on the holdout set is worse than what you report. There also appear to be multiple evaluation systems you have set up, which is the best to use to replicate your best results?

Thanks - Dennis

After the full dataset is released, we will add the usage of the code.