How to use TTA for mmrotate1.x?

[Haru-zt]

Hi, When I use the TTA in mmdet for the test in mmrotate1.x, I found that the function merge_aug_bboxes need a input tensor like aug_bboxes (list[Tensor]): shape (n, 4#class) .
However, for now, mmrotate1.x only support the output fomat like (x, y, w, h, theta), so i have no way to convert it.
Is there any tips for me to use TTA in mmrotate1.x correctly? Maybe I miss some document.
Thanks for help.

[Haru-zt]

Hi, I use the change in open-mmlab/mmrotate#771.
tta works, but the result is bad. I don't know why tta produce a lot of wrong bboxes. I don't know why.

[hhaAndroid]

@Haru-zt Please to mmrotate create a issue