NonMaxSuppression#

NonMaxSuppression - 11#

Version

  • name: NonMaxSuppression (GitHub)

  • domain: main

  • since_version: 11

  • function: False

  • support_level: SupportType.COMMON

  • shape inference: True

This version of the operator has been available since version 11.

Summary

Filter out boxes that have high intersection-over-union (IOU) overlap with previously selected boxes. Bounding boxes with score less than score_threshold are removed. Bounding box format is indicated by attribute center_point_box. Note that this algorithm is agnostic to where the origin is in the coordinate system and more generally is invariant to orthogonal transformations and translations of the coordinate system; thus translating or reflections of the coordinate system result in the same boxes being selected by the algorithm. The selected_indices output is a set of integers indexing into the input collection of bounding boxes representing the selected boxes. The bounding box coordinates corresponding to the selected indices can then be obtained using the Gather or GatherND operation.

Attributes

  • center_point_box: Integer indicate the format of the box data. The default is 0. 0 - the box data is supplied as [y1, x1, y2, x2] where (y1, x1) and (y2, x2) are the coordinates of any diagonal pair of box corners and the coordinates can be provided as normalized (i.e., lying in the interval [0, 1]) or absolute. Mostly used for TF models. 1 - the box data is supplied as [x_center, y_center, width, height]. Mostly used for Pytorch models. Default value is 0.

Inputs

Between 2 and 5 inputs.

  • boxes (heterogeneous) - tensor(float): An input tensor with shape [num_batches, spatial_dimension, 4]. The single box data format is indicated by center_point_box.

  • scores (heterogeneous) - tensor(float): An input tensor with shape [num_batches, num_classes, spatial_dimension]

  • max_output_boxes_per_class (optional, heterogeneous) - tensor(int64): Integer representing the maximum number of boxes to be selected per batch per class. It is a scalar. Default to 0, which means no output.

  • iou_threshold (optional, heterogeneous) - tensor(float): Float representing the threshold for deciding whether boxes overlap too much with respect to IOU. It is scalar. Value range [0, 1]. Default to 0.

  • score_threshold (optional, heterogeneous) - tensor(float): Float representing the threshold for deciding when to remove boxes based on score. It is a scalar.

Outputs

  • selected_indices (heterogeneous) - tensor(int64): selected indices from the boxes tensor. [num_selected_indices, 3], the selected index format is [batch_index, class_index, box_index].

Examples

nonmaxsuppression_suppress_by_IOU

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.1, 1.0, 1.1],
    [0.0, -0.1, 1.0, 0.9],
    [0.0, 10.0, 1.0, 11.0],
    [0.0, 10.1, 1.0, 11.1],
    [0.0, 100.0, 1.0, 101.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0], [0, 0, 5]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_suppress_by_IOU')

nonmaxsuppression_suppress_by_IOU_and_scores

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.1, 1.0, 1.1],
    [0.0, -0.1, 1.0, 0.9],
    [0.0, 10.0, 1.0, 11.0],
    [0.0, 10.1, 1.0, 11.1],
    [0.0, 100.0, 1.0, 101.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.4]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_suppress_by_IOU_and_scores')

nonmaxsuppression_flipped_coordinates

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [1.0, 1.0, 0.0, 0.0],
    [0.0, 0.1, 1.0, 1.1],
    [0.0, 0.9, 1.0, -0.1],
    [0.0, 10.0, 1.0, 11.0],
    [1.0, 10.1, 0.0, 11.1],
    [1.0, 101.0, 0.0, 100.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0], [0, 0, 5]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_flipped_coordinates')

nonmaxsuppression_limit_output_size

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.1, 1.0, 1.1],
    [0.0, -0.1, 1.0, 0.9],
    [0.0, 10.0, 1.0, 11.0],
    [0.0, 10.1, 1.0, 11.1],
    [0.0, 100.0, 1.0, 101.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([2]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_limit_output_size')

nonmaxsuppression_single_box

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0]
]]).astype(np.float32)
scores = np.array([[[0.9]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_single_box')

nonmaxsuppression_identical_boxes

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],

    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.0, 1.0, 1.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_identical_boxes')

nonmaxsuppression_center_point_box_format

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices'],
    center_point_box=1
)
boxes = np.array([[
    [0.5, 0.5, 1.0, 1.0],
    [0.5, 0.6, 1.0, 1.0],
    [0.5, 0.4, 1.0, 1.0],
    [0.5, 10.5, 1.0, 1.0],
    [0.5, 10.6, 1.0, 1.0],
    [0.5, 100.5, 1.0, 1.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([3]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0], [0, 0, 5]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_center_point_box_format')

nonmaxsuppression_two_classes

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[
    [0.0, 0.0, 1.0, 1.0],
    [0.0, 0.1, 1.0, 1.1],
    [0.0, -0.1, 1.0, 0.9],
    [0.0, 10.0, 1.0, 11.0],
    [0.0, 10.1, 1.0, 11.1],
    [0.0, 100.0, 1.0, 101.0]
]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3],
                    [0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([2]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0], [0, 1, 3], [0, 1, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_two_classes')

nonmaxsuppression_two_batches

node = onnx.helper.make_node(
    'NonMaxSuppression',
    inputs=['boxes', 'scores', 'max_output_boxes_per_class', 'iou_threshold', 'score_threshold'],
    outputs=['selected_indices']
)
boxes = np.array([[[0.0, 0.0, 1.0, 1.0],
                   [0.0, 0.1, 1.0, 1.1],
                   [0.0, -0.1, 1.0, 0.9],
                   [0.0, 10.0, 1.0, 11.0],
                   [0.0, 10.1, 1.0, 11.1],
                   [0.0, 100.0, 1.0, 101.0]],
                  [[0.0, 0.0, 1.0, 1.0],
                   [0.0, 0.1, 1.0, 1.1],
                   [0.0, -0.1, 1.0, 0.9],
                   [0.0, 10.0, 1.0, 11.0],
                   [0.0, 10.1, 1.0, 11.1],
                   [0.0, 100.0, 1.0, 101.0]]]).astype(np.float32)
scores = np.array([[[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]],
                   [[0.9, 0.75, 0.6, 0.95, 0.5, 0.3]]]).astype(np.float32)
max_output_boxes_per_class = np.array([2]).astype(np.int64)
iou_threshold = np.array([0.5]).astype(np.float32)
score_threshold = np.array([0.0]).astype(np.float32)
selected_indices = np.array([[0, 0, 3], [0, 0, 0], [1, 0, 3], [1, 0, 0]]).astype(np.int64)

expect(node, inputs=[boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold], outputs=[selected_indices], name='test_nonmaxsuppression_two_batches')

Differences

00Filter out boxes that have high intersection-over-union (IOU) overlap with previously selected boxes.Filter out boxes that have high intersection-over-union (IOU) overlap with previously selected boxes.
11Bounding boxes with score less than score_threshold are removed. Bounding box format is indicated by attribute center_point_box.Bounding boxes with score less than score_threshold are removed. Bounding box format is indicated by attribute center_point_box.
22Note that this algorithm is agnostic to where the origin is in the coordinate system and more generally is invariant toNote that this algorithm is agnostic to where the origin is in the coordinate system and more generally is invariant to
33orthogonal transformations and translations of the coordinate system; thus translating or reflections of the coordinate systemorthogonal transformations and translations of the coordinate system; thus translating or reflections of the coordinate system
44result in the same boxes being selected by the algorithm.result in the same boxes being selected by the algorithm.
55The selected_indices output is a set of integers indexing into the input collection of bounding boxes representing the selected boxes.The selected_indices output is a set of integers indexing into the input collection of bounding boxes representing the selected boxes.
66The bounding box coordinates corresponding to the selected indices can then be obtained using the Gather or GatherND operation.The bounding box coordinates corresponding to the selected indices can then be obtained using the Gather or GatherND operation.
77
88**Attributes****Attributes**
99
1010* **center_point_box**:* **center_point_box**:
1111 Integer indicate the format of the box data. The default is 0. 0 - Integer indicate the format of the box data. The default is 0. 0 -
1212 the box data is supplied as [y1, x1, y2, x2] where (y1, x1) and (y2, the box data is supplied as [y1, x1, y2, x2] where (y1, x1) and (y2,
1313 x2) are the coordinates of any diagonal pair of box corners and the x2) are the coordinates of any diagonal pair of box corners and the
1414 coordinates can be provided as normalized (i.e., lying in the coordinates can be provided as normalized (i.e., lying in the
1515 interval [0, 1]) or absolute. Mostly used for TF models. 1 - the box interval [0, 1]) or absolute. Mostly used for TF models. 1 - the box
1616 data is supplied as [x_center, y_center, width, height]. Mostly used data is supplied as [x_center, y_center, width, height]. Mostly used
1717 for Pytorch models. Default value is 0. for Pytorch models. Default value is 0.
1818
1919**Inputs****Inputs**
2020
2121Between 2 and 5 inputs.Between 2 and 5 inputs.
2222
2323* **boxes** (heterogeneous) - **tensor(float)**:* **boxes** (heterogeneous) - **tensor(float)**:
2424 An input tensor with shape [num_batches, spatial_dimension, 4]. The An input tensor with shape [num_batches, spatial_dimension, 4]. The
2525 single box data format is indicated by center_point_box. single box data format is indicated by center_point_box.
2626* **scores** (heterogeneous) - **tensor(float)**:* **scores** (heterogeneous) - **tensor(float)**:
2727 An input tensor with shape [num_batches, num_classes, An input tensor with shape [num_batches, num_classes,
2828 spatial_dimension] spatial_dimension]
2929* **max_output_boxes_per_class** (optional, heterogeneous) - **tensor(int64)**:* **max_output_boxes_per_class** (optional, heterogeneous) - **tensor(int64)**:
3030 Integer representing the maximum number of boxes to be selected per Integer representing the maximum number of boxes to be selected per
3131 batch per class. It is a scalar. Default to 0, which means no batch per class. It is a scalar. Default to 0, which means no
3232 output. output.
3333* **iou_threshold** (optional, heterogeneous) - **tensor(float)**:* **iou_threshold** (optional, heterogeneous) - **tensor(float)**:
3434 Float representing the threshold for deciding whether boxes overlap Float representing the threshold for deciding whether boxes overlap
3535 too much with respect to IOU. It is scalar. Value range [0, 1]. too much with respect to IOU. It is scalar. Value range [0, 1].
3636 Default to 0. Default to 0.
3737* **score_threshold** (optional, heterogeneous) - **tensor(float)**:* **score_threshold** (optional, heterogeneous) - **tensor(float)**:
3838 Float representing the threshold for deciding when to remove boxes Float representing the threshold for deciding when to remove boxes
3939 based on score. It is a scalar. based on score. It is a scalar.
4040
4141**Outputs****Outputs**
4242
4343* **selected_indices** (heterogeneous) - **tensor(int64)**:* **selected_indices** (heterogeneous) - **tensor(int64)**:
4444 selected indices from the boxes tensor. [num_selected_indices, 3], selected indices from the boxes tensor. [num_selected_indices, 3],
4545 the selected index format is [batch_index, class_index, box_index]. the selected index format is [batch_index, class_index, box_index].

NonMaxSuppression - 10#

Version

  • name: NonMaxSuppression (GitHub)

  • domain: main

  • since_version: 10

  • function: False

  • support_level: SupportType.COMMON

  • shape inference: True

This version of the operator has been available since version 10.

Summary

Filter out boxes that have high intersection-over-union (IOU) overlap with previously selected boxes. Bounding boxes with score less than score_threshold are removed. Bounding box format is indicated by attribute center_point_box. Note that this algorithm is agnostic to where the origin is in the coordinate system and more generally is invariant to orthogonal transformations and translations of the coordinate system; thus translating or reflections of the coordinate system result in the same boxes being selected by the algorithm. The selected_indices output is a set of integers indexing into the input collection of bounding boxes representing the selected boxes. The bounding box coordinates corresponding to the selected indices can then be obtained using the Gather or GatherND operation.

Attributes

  • center_point_box: Integer indicate the format of the box data. The default is 0. 0 - the box data is supplied as [y1, x1, y2, x2] where (y1, x1) and (y2, x2) are the coordinates of any diagonal pair of box corners and the coordinates can be provided as normalized (i.e., lying in the interval [0, 1]) or absolute. Mostly used for TF models. 1 - the box data is supplied as [x_center, y_center, width, height]. Mostly used for Pytorch models. Default value is 0.

Inputs

Between 2 and 5 inputs.

  • boxes (heterogeneous) - tensor(float): An input tensor with shape [num_batches, spatial_dimension, 4]. The single box data format is indicated by center_point_box.

  • scores (heterogeneous) - tensor(float): An input tensor with shape [num_batches, num_classes, spatial_dimension]

  • max_output_boxes_per_class (optional, heterogeneous) - tensor(int64): Integer representing the maximum number of boxes to be selected per batch per class. It is a scalar. Default to 0, which means no output.

  • iou_threshold (optional, heterogeneous) - tensor(float): Float representing the threshold for deciding whether boxes overlap too much with respect to IOU. It is scalar. Value range [0, 1]. Default to 0.

  • score_threshold (optional, heterogeneous) - tensor(float): Float representing the threshold for deciding when to remove boxes based on score. It is a scalar.

Outputs

  • selected_indices (heterogeneous) - tensor(int64): selected indices from the boxes tensor. [num_selected_indices, 3], the selected index format is [batch_index, class_index, box_index].