What does hbox rbox qbox mean in mmrotate config?

[vansin]

mmrotate的config中的 hbox 、rbox和qbox是什么意思？有什么区别吗？
What does hbox mean?

[vansin]

对比以下两个的文件可以发现，hbox 和 qbox 最大的区别只是 assigner中的iou_calculator从RBboxOverlaps2D变成了FakeRBboxOverlaps2D
https://github.com/open-mmlab/mmrotate/blob/dev-1.x/configs/rotated_retinanet/rotated-retinanet-rbox-le90_r50_fpn_1x_dota.py
https://github.com/open-mmlab/mmrotate/blob/dev-1.x/configs/rotated_retinanet/rotated-retinanet-hbox-le90_r50_fpn_1x_dota.py

_base_ = './rotated-retinanet-rbox-le90_r50_fpn_1x_dota.py'

model = dict(
    train_cfg=dict(
        assigner=dict(iou_calculator=dict(type='FakeRBboxOverlaps2D'))))

train_pipeline = [
    dict(
        type='mmdet.LoadImageFromFile',
        file_client_args={{_base_.file_client_args}}),
    dict(type='mmdet.LoadAnnotations', with_bbox=True, box_type='qbox'),
    dict(type='ConvertBoxType', box_type_mapping=dict(gt_bboxes='rbox')),
    dict(type='mmdet.Resize', scale=(1024, 1024), keep_ratio=True),
    dict(type='mmdet.RandomFlip', prob=0.5),
    dict(type='mmdet.PackDetInputs')
]

train_dataloader = dict(dataset=dict(pipeline=train_pipeline))

[vansin]

从以下代码段中的 Calculate IoU between 2D minimum circumscribed hbbs of rbbs 这句话，搞懂了 hbox 在 config 中的意思，是训练的时候计算iou是通过求最小外接水平矩形，然后两个水平矩形再求iou。
而rbox是旋转框直接求iou，引申的疑问是为什么要设置这个对比试验？

@TASK_UTILS.register_module()
class FakeRBboxOverlaps2D(object):
    """2D Overlaps Calculator for Minimum Circumscribed Horizental Bboxes of
    Rotated Bboxes."""

    def __call__(self,
                 bboxes1: RotatedBoxes,
                 bboxes2: RotatedBoxes,
                 mode: str = 'iou',
                 is_aligned: bool = False) -> Tensor:
        """Calculate IoU between 2D minimum circumscribed hbbs of rbbs.

        Args:
            bboxes1 (:obj:`RotatedBoxes` or Tensor): bboxes have shape (m, 5)
                in <cx, cy, w, h, t> format, shape (m, 6) in
                <cx, cy, w, h, t, score> format.
            bboxes2 (:obj:`RotatedBoxes` or Tensor): bboxes have shape (n, 5)
                in <cx, cy, w, h, t> format, shape (n, 6) in
                <cx, cy, w, h, t, score> format, or be empty.
            mode (str): 'iou' (intersection over union), 'iof' (intersection
                over foreground).
            is_aligned (bool): If True, then m and n must be equal.
                Defaults to False.

        Returns:
            Tensor: shape (m, n) if ``is_aligned `` is False else shape (m,)
        """
        assert bboxes1.size(-1) in [0, 5, 6]
        assert bboxes2.size(-1) in [0, 5, 6]

        if bboxes1.size(-1) == 6:
            bboxes1 = bboxes1[..., :5]
        if bboxes2.size(-1) == 6:
            bboxes2 = bboxes2[..., :5]

        if not isinstance(bboxes1, RotatedBoxes):
            bboxes1 = RotatedBoxes(bboxes1)
        if not isinstance(bboxes2, RotatedBoxes):
            bboxes2 = RotatedBoxes(bboxes2)

        return fake_rbbox_overlaps(bboxes1, bboxes2, mode, is_aligned)

    def __repr__(self) -> str:
        """str: a string describing the module"""
        repr_str = self.__class__.__name__ + '()'
        return repr_str

[jamiechoi1995]

是不是因为rbox更精细，所以mAP更高？

mmrotate/configs/rotated_retinanet/README.md

Lines 21 to 22 in b55be23

| ResNet50 (1024,1024,200) | 64.55 | oc | 1x | 3.38 | 15.7 | - | 2 | [rotated-retinanet-hbox-oc_r50_fpn_1x_dota](./rotated-retinanet-hbox-oc_r50_fpn_1x_dota.py) | [model](https://download.openmmlab.com/mmrotate/v0.1.0/rotated_retinanet/rotated_retinanet_hbb_r50_fpn_1x_dota_oc/rotated_retinanet_hbb_r50_fpn_1x_dota_oc-e8a7c7df.pth) \| [log](https://download.openmmlab.com/mmrotate/v0.1.0/rotated_retinanet/rotated_retinanet_hbb_r50_fpn_1x_dota_oc/rotated_retinanet_hbb_r50_fpn_1x_dota_oc_20220121_095315.log.json) |

| ResNet50 (1024,1024,200) | 68.42 | le90 | 1x | 3.38 | 16.9 | - | 2 | [rotated-retinanet-rbox-le90_r50_fpn_1x_dota](./rotated-retinanet-rbox-le90_r50_fpn_1x_dota.py) | [model](https://download.openmmlab.com/mmrotate/v0.1.0/rotated_retinanet/rotated_retinanet_obb_r50_fpn_1x_dota_le90/rotated_retinanet_obb_r50_fpn_1x_dota_le90-c0097bc4.pth) \| [log](https://download.openmmlab.com/mmrotate/v0.1.0/rotated_retinanet/rotated_retinanet_obb_r50_fpn_1x_dota_le90/rotated_retinanet_obb_r50_fpn_1x_dota_le90_20220128_130740.log.json) |