Support the instance segmentation task in projects

[zytx121]

This project implements a Mask RCNN for rotated boxes. Benefiting from the BoxType design, we only need to modify the code slightly in mmrotate, and we can support the instance segmentation task. Although the rotated box indicator will drop slightly after adding the mask head, it may help improve the instance segmentation task.

We hope this project can inspire you and welcome you to explore more uses of mmrotate!

[RangiLyu]

Do you have model weights to upload?

[fkroeber]

Thanks for this contribution, I believe it's very valuable to expand the libraries functionality!
One suggestion for an improvement of the mask delineation: Replace the proposed matching of rotated ROI aligned feature maps with hbox mask targets (using rbox2hox in mask_target.py -> mask_target_single & rotated_fcn_mask_head.py -> _do_paste_mask()) with an rotated ROI alignment for the mask targets. Thereby the feature maps as well as the mask targets are actually spatially matching, which allows a better mask prediction. To illustrate the improvement, you can find a comparison of both approaches for the use case of agricultural parcel delineation below.

[yangxue0827]

@zytx121

[Ataraxi8]

When I use my own training set, why the box is shifted, but the segm is correct? @zytx121 @fkroeber

[zytx121]

@fkroeber Thank you for sharing, the results are much better than now! May I ask if you are willing to provide us with a PR?

[tcotte]

Hello,
I am new to MMRotate. I've found this library while I was searching for an instance segmentation model with oriented bounding boxes. Thank you @zytx121, I think this project will solve my problem.
I have seen that DOTA dataset type is privileged in MMRotate. But I guess that masks have to be provided. Therefore, I assume that DOTA dataset is not supported by your project. This is why I ask you what type of dataset have to be used in this project ?