result

[2276924877]

Hello author, I used mmrotate for training and testing. https://github.com/open-mmlab/mmrotateI encountered some issues with the test results, as shown in the figure below, where there are some regression boxes that do not fully enclose the target. I wanted to ask if you also encountered similar issues during your experiments, or if there were any mistakes I made during training.

Enviroment：ubuntu18.04
python 3.8
cuda 11.1
torch 1.9.1
mmcv 1.7.1
mmrotate 0.3.4

Config：orinted_reppoints_r50_fpn_1x_dota_le135