ExperimentalDetectronDetectionOutput

Versioned name: ExperimentalDetectronDetectionOutput-6

Category: Object detection

Short description: The ExperimentalDetectronDetectionOutput operation performs non-maximum suppression to generate the detection output using information on location and score predictions.

Detailed description: The operation performs the following steps:

1. Applies deltas to boxes sizes [x 1, y 1, x 2, y 2] and takes coordinates of refined boxes according to the formulas:

x1_new = ctr_x + (dx - 0.5 * exp(min(d_log_w, max_delta_log_wh))) * box_w

y0_new = ctr_y + (dy - 0.5 * exp(min(d_log_h, max_delta_log_wh))) * box_h

x1_new = ctr_x + (dx + 0.5 * exp(min(d_log_w, max_delta_log_wh))) * box_w - 1.0

y1_new = ctr_y + (dy + 0.5 * exp(min(d_log_h, max_delta_log_wh))) * box_h - 1.0

• box_w and box_h are width and height of box, respectively:

box_w = x1 - x0 + 1.0

box_h = y1 - y0 + 1.0

• ctr_x and ctr_y are center location of a box:

ctr_x = x0 + 0.5f * box_w

ctr_y = y0 + 0.5f * box_h

• dx, dy, d_log_w and d_log_h are deltas calculated according to the formulas below, and deltas_tensor is a second input:

dx = deltas_tensor[roi_idx, 4 * class_idx + 0] / deltas_weights[0]

dy = deltas_tensor[roi_idx, 4 * class_idx + 1] / deltas_weights[1]

d_log_w = deltas_tensor[roi_idx, 4 * class_idx + 2] / deltas_weights[2]

d_log_h = deltas_tensor[roi_idx, 4 * class_idx + 3] / deltas_weights[3]

2. If class_agnostic_box_regression is true removes predictions for background classes.

3. Clips boxes to the image.

4. Applies score_threshold on detection scores.

5. Applies non-maximum suppression class-wise with nms_threshold and returns post_nms_count or less detections per class.

6. Returns max_detections_per_image detections if total number of detections is more than max_detections_per_image; otherwise, returns total number of detections and the output tensor is filled with undefined values for rest output tensor elements.

Attributes:

• score_threshold

  • Description: The score_threshold attribute specifies a threshold to consider only detections whose score are larger than the threshold.

  • Range of values: non-negative floating-point number

  • Type: float

  • Default value: None

  • Required: yes

• nms_threshold

  • Description: The nms_threshold attribute specifies a threshold to be used in the NMS stage.

  • Range of values: non-negative floating-point number

  • Type: float

  • Default value: None

  • Required: yes

• num_classes

  • Description: The num_classes attribute specifies the number of detected classes.

  • Range of values: non-negative integer number

  • Type: int

  • Default value: None

  • Required: yes

• post_nms_count

  • Description: The post_nms_count attribute specifies the maximal number of detections per class.

  • Range of values: non-negative integer number

  • Type: int

  • Default value: None

  • Required: yes

• max_detections_per_image

  • Description: The max_detections_per_image attribute specifies maximal number of detections per image.

  • Range of values: non-negative integer number

  • Type: int

  • Default value: None

  • Required: yes

• class_agnostic_box_regression

  • Description: class_agnostic_box_regression attribute is a flag that specifies whether to delete background classes or not.

  • Range of values:

    • true means background classes should be deleted

    • false means background classes should not be deleted

  • Type: boolean

  • Default value: false

  • Required: no

• max_delta_log_wh

  • Description: The max_delta_log_wh attribute specifies maximal delta of logarithms for width and height.

  • Range of values: floating-point number

  • Type: float

  • Default value: None

  • Required: yes

• deltas_weights

  • Description: The deltas_weights attribute specifies weights for bounding boxes sizes deltas.

  • Range of values: a list of non-negative floating-point numbers

  • Type: float[]

  • Default value: None

  • Required: yes

Inputs

• 1: A 2D tensor of type T with input ROIs, with shape [number_of_ROIs, 4] providing the ROIs as 4-tuples: [x 1, y 1, x 2, y 2]. The batch dimension of first, second, and third inputs should be the same. Required.

• 2: A 2D tensor of type T with shape [number_of_ROIs, num_classes * 4] providing deltas for input boxes. Required.

• 3: A 2D tensor of type T with shape [number_of_ROIs, num_classes] providing detections scores. Required.

• 4: A 2D tensor of type T with shape [1, 3] contains three elements [image_height, image_width, scale_height_and_width] providing input image size info. Required.

Outputs

• 1: A 2D tensor of type T with shape [max_detections_per_image, 4] providing boxes indices.

• 2: A 1D tensor of type T_IND with shape [max_detections_per_image] providing classes indices.

• 3: A 1D tensor of type T with shape [max_detections_per_image] providing scores indices.

Types

• T: any supported floating-point type.

• T_IND: int64 or int32.

Example

<layer ... type="ExperimentalDetectronDetectionOutput" version="opset6">
    <data class_agnostic_box_regression="false" deltas_weights="10.0,10.0,5.0,5.0" max_delta_log_wh="4.135166645050049" max_detections_per_image="100" nms_threshold="0.5" num_classes="81" post_nms_count="2000" score_threshold="0.05000000074505806"/>
    <input>
        <port id="0">
            <dim>1000</dim>
            <dim>4</dim>
        </port>
        <port id="1">
            <dim>1000</dim>
            <dim>324</dim>
        </port>
        <port id="2">
            <dim>1000</dim>
            <dim>81</dim>
        </port>
        <port id="3">
            <dim>1</dim>
            <dim>3</dim>
        </port>
    </input>
    <output>
        <port id="4" precision="FP32">
            <dim>100</dim>
            <dim>4</dim>
        </port>
        <port id="5" precision="I32">
            <dim>100</dim>
        </port>
        <port id="6" precision="FP32">
            <dim>100</dim>
        </port>
        <port id="7" precision="I32">
            <dim>100</dim>
        </port>
    </output>
</layer>