Performance on DOTA_v1.0 val set

[TianhangTang]

Hi, I downloaded the R2CNN trained weight of DOTA (which is "FasterRCNN_20180515_DOTA_v3") and DOTA dataset, I just use the val_crop.py to generate val set and do evaluate test on it, I got both oriented test and horizontal test map value about 37%, the result turns out to be much worse than the given result which is 60.37%.
I wonder is there anything wrong on my test operation, or the given trained weights is not the final converged weights, or the DOTA dataset has been greatly changed cause DOTA_v1.5 has been released even though I am using the label_v1.0?
Thank you