mindspore.ops.NMSWithMask

class mindspore.ops.NMSWithMask(*args, **kwargs)[source]

When object detection problem is performed in the computer vision field, object detection algorithm generates a plurality of bounding boxes. Selects some bounding boxes in descending order of score(Descending order is not supported in Ascend platform currently). Use the box with the highest score calculate the overlap between other boxes and the current box, and delete the box based on a certain threshold(IOU). The IOU is as follows,

\[\text{IOU} = \frac{\text{Area of Overlap}}{\text{Area of Union}}\]
Parameters

iou_threshold (float) – Specifies the threshold of overlap boxes with respect to IOU. Default: 0.5.

Inputs:

  • bboxes (Tensor) - The shape of tensor is \((N, 5)\). Input bounding boxes. N is the number of input bounding boxes. Every bounding box contains 5 values, the first 4 values are the coordinates(x0, y0, x1, y1) of bounding box which represents the point of top-left and bottom-right, and the last value is the score of this bounding box. The data type must be float16 or float32.

Outputs:

tuple[Tensor], tuple of three tensors, they are selected_boxes, selected_idx and selected_mask.

  • selected_boxes (Tensor) - The shape of tensor is \((N, 5)\). The list of bounding boxes after non-max suppression calculation.

  • selected_idx (Tensor) - The shape of tensor is \((N,)\). The indexes list of valid input bounding boxes.

  • selected_mask (Tensor) - The shape of tensor is \((N,)\). A mask list of valid output bounding boxes.

Raises

ValueError – If the iou_threshold is not a float number, or if the first dimension of input Tensor is less than or equal to 0, or if the data type of the input Tensor is not float16 or float32.

Supported Platforms:

Ascend GPU

Examples

>>> bbox = np.array([[100.0, 100.0, 50.0, 68.0, 0.63], [150.0, 75.0, 165.0, 115.0, 0.55],
...                  [12.0, 190.0, 288.0, 200.0, 0.9], [28.0, 130.0, 106.0, 172.0, 0.3]])
>>> bbox[:, 2] += bbox[:, 0]
>>> bbox[:, 3] += bbox[:, 1]
>>> inputs = Tensor(bbox, mindspore.float32)
>>> nms = ops.NMSWithMask(0.1)
>>> output_boxes, indices, mask = nms(inputs)
>>> indices_np = indices.asnumpy()
>>> print(indices_np[mask.asnumpy()])
[0 1 2]