mindspore.ops.rotated_iou

View Source On Gitee
mindspore.ops.rotated_iou(boxes, query_boxes, trans=False, mode=0, is_cross=True, v_threshold=0.0, e_threshold=0.0)[source]

Calculate the overlap area between rotated rectangles.

Warning

This is an experimental API that is subject to change or deletion.

Note

The input data types supported by the Ascend platform include bfloat16, float16, float32.

Parameters

  • boxes (Tensor) – The first set of rectangles which has a shape of \((B, N, 5)\).

  • query_boxes (Tensor) – The second set of rectangles which has a shape of \((B, K, 5)\).

  • trans (bool) – Distinguish the rectangles representations of boxes and query_boxes. If True, the format of boxes and query_boxes is 'xyxyt', else the format is 'xywht'. The default value is False.

  • mode (int) – Distinguish the calculation mode. If the value is 1, the calculation mode is 'iof', else the calculation mode is 'iou'. The default value is 0.

  • is_cross (bool) – If True, use cross-calculation, else use one-to-one calculation. The default value is True.

  • v_threshold (float) – Provide condition relaxation for intersection calculation. The default value is 0.0.

  • e_threshold (float) – Provide condition relaxation for intersection calculation. The default value is 0.0.

Returns

Tensor, the shape is \((B, N, K)\).

Raises

  • TypeError – If boxes is not a Tensor.

  • TypeError – If query_boxes is not a Tensor.

  • ValueError – If boxes and query_boxes do not has same first dim.

  • ValueError – If the third dimension of boxes or query_boxes is not 5.

Supported Platforms:

Ascend

Examples

>>> import mindspore
>>> import numpy as np
>>> from mindspore import Tensor, ops
>>> a = np.random.uniform(0,1,(2,2,5)).astype(np.float16)
>>> b = np.random.uniform(0,1,(2,3,5)).astype(np.float16)
>>> box1 = Tensor(a)
>>> box2 = Tensor(b)
>>> output = ops.rotated_iou(box1, box2, trans=False, mode=0, is_cross=True)