[Feature] Add points_in_polygons CUDA op for rotated detection.

[zytx121]

Add points_in_polygons CUDA ops for anchor-free rotated detection. It is mainly used to judge whether the points are inside the ground true polygons during sample assignment.

Modification

Several files in mmcv/ops folder.

BC-breaking (Optional)

No.

Use cases (Optional)

from mmcv.ops import points_in_polygons

[zhouzaida]

Please @AllentDan have a look.

[AllentDan]

Hi, @zytx121. Unit tests failed on my device.

[zytx121]

Hi, @zytx121. Unit tests failed on my device.

Sorry, I'll check it as soon as possible.

[zytx121]

Hi, @zytx121. Unit tests failed on my device.

Please test again. Thx.

[AllentDan]

LGTM except updating ops in https://github.com/open-mmlab/mmcv/blob/master/docs/en/understand_mmcv/ops.md and https://github.com/open-mmlab/mmcv/blob/master/docs/zh_cn/understand_mmcv/ops.md

[grimoire]

Is there any limit about the polygon? Such as convex or clockwise?

Test failed with data:

    points = np.array([[20, 90]])
    polygons = np.array([[0., 0., 100., 100., 10., 90., 0., 100.]])

If there are any limit about the polygon, please update the docstring.

[zytx121]

Is there any limit about the polygon? Such as convex or clockwise?

Test failed with data:

    points = np.array([[20, 90]])
    polygons = np.array([[0., 0., 100., 100., 10., 90., 0., 100.]])

If there are any limit about the polygon, please update the docstring.

We believe that this method can handle all polygons without adding conditions such as convex or clockwise. As shown in the figure below, this point falls inside the polygon, which is consistent with the test results.

def test_points_in_polygons():
    points = np.array([[20, 90]])
    polygons = np.array([[0., 0., 100., 100., 10., 90., 0., 100.]])
    expert_output = np.array([[1.,]])
    points = torch.from_numpy(points).cuda().float()
    polygons = torch.from_numpy(polygons).cuda().float()
    expert_output = torch.from_numpy(expert_output).cuda().float()
    assert torch.allclose(
        points_in_polygons(points, polygons), expert_output, 1e-3)

Note that the format of the polygon is (x,y,x,y,x,y,x,y).

[grimoire]

@zytx121 Thanks for the explanation. The data did pass the test as you said.

[grimoire]

LGTM.

[ZwwWayne]

Can be merged after resolving conflicts.