OrientedRepPoints config with ms in MMRotate seems not to use ms RResize transformation

[songyuc]

Hi, guys,
I am learning the OrientedRepPoints model in MMRotate. But I find the oriented_reppoints_r50_fpn_40e_dota_ms_le135.py config seems not to use ms RResize transformation, when training with command of python tools/misc/browse_dataset.py configs/oriented_reppoints/oriented_reppoints_r50_fpn_40e_dota_ms_le135.py. I debugged it in PyCharm, and find its pipeline is shown like:

which shows it uses the img_scale=(1024, 1024) RResize, instead of the img_scale=[(1333, 768), (1333, 1280)] RResize.

Could you please offer some guide?
Your answer and explanation will be appreciated!

[yangxue0827]

MS in MMRotate is offline, you should use MS config when preparing datasets (crop DOTA dataset).

[songyuc]

Hi, @yangxue0827,
I have used [split dota dataset] in MMRotate, however, which does not require a model config as input.
So, what do you mean with "use MS config when preparing datasets (crop DOTA dataset)"?

[yangxue0827]

ms means multi-scale crop. It is different from the common multi-scale online training (e.g. img_scale=[(1333, 768), (1333, 1280)]), but the purpose is the same.

[songyuc]

Thanks sincerely for your explanation!