DOTA split #7
Description
Hello, thank you for your great work!
I'm trying to reproduce the dataset splitting results using the official DOTA v1.0 image_split tool with subsize = 512 and gap = 100. After splitting the DOTA v1.0 train set, I obtained 59,753 sub-images in total. After removing images with no annotations, I got 24,557 valid samples.
However, this number is significantly higher than the 17,609 training samples reported in your paper/documentation.
Could you please let us know if any post-processing steps were applied to achieve the reported number of samples, and if so, could you provide details on how these steps were conducted?
Thank you again!