Skip to content

Latest commit

 

History

History
140 lines (99 loc) · 6.88 KB

ns-directml-dml_roi_align_operator_desc.md

File metadata and controls

140 lines (99 loc) · 6.88 KB
Raw
UID title description helpviewer_keywords tech.root ms.date req.header req.include-header req.target-type req.target-min-winverclnt req.target-min-winversvr req.kmdf-ver req.umdf-ver req.ddi-compliance req.unicode-ansi req.idl req.max-support req.namespace req.assembly req.type-library req.lib req.dll req.irql targetos req.typenames req.redist f1_keywords dev_langs topic_type api_type api_location api_name
NS:directml.DML_ROI_ALIGN_OPERATOR_DESC
DML_ROI_ALIGN_OPERATOR_DESC
Performs an ROI align operation, as described in the [Mask R-CNN](https://arxiv.org/abs/1703.06870) paper. In summary, the operation extracts crops from the input image tensor and resizes them to a common output size specified by the last 2 dimensions of *OutputTensor* using the specified *InterpolationMode*.
DML_ROI_ALIGN_OPERATOR_DESC
DML_ROI_ALIGN_OPERATOR_DESC structure
direct3d12.dml_roi_align_operator_desc
directml/DML_ROI_ALIGN_OPERATOR_DESC
directml
01/19/2022
directml.h
Windows
Windows 10 Build 20348
Windows 10 Build 20348
Windows
DML_ROI_ALIGN_OPERATOR_DESC
directml/DML_ROI_ALIGN_OPERATOR_DESC
c++
APIRef
kbSyntax
HeaderDef
DirectML.h
DML_ROI_ALIGN_OPERATOR_DESC

-description

Performs an ROI align operation, as described in the Mask R-CNN paper. In summary, the operation extracts crops from the input image tensor and resizes them to a common output size specified by the last 2 dimensions of OutputTensor using the specified InterpolationMode.

-struct-fields

-field InputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the input data with dimensions { BatchCount, ChannelCount, InputHeight, InputWidth }.

-field ROITensor

Type: const DML_TENSOR_DESC*

A tensor containing the regions of interest (ROI) data. The allowed dimensions of ROITensor are { NumROIs, 4 }, { 1, NumROIs, 4 }, or { 1, 1, NumROIs, 4 }. For each ROI, the values will be the coordinates of its top-left and bottom-right corners in the order [x1, y1, x2, y2].

-field BatchIndicesTensor

Type: const DML_TENSOR_DESC*

A tensor containing the batch indices to extract the ROIs from. The allowed dimensions of BatchIndicesTensor are { NumROIs }, { 1, NumROIs }, { 1, 1, NumROIs }, or { 1, 1, 1, NumROIs }. Each value is the index of a batch from InputTensor. The behavior is undefined if the values are not in the range [0, BatchCount).

-field OutputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the output data. The expected dimensions of OutputTensor are { NumROIs, ChannelCount, OutputHeight, OutputWidth }.

-field ReductionFunction

Type: DML_REDUCE_FUNCTION

The reduction function to use when reducing across all input samples that contribute to an output element (DML_REDUCE_FUNCTION_AVERAGE or DML_REDUCE_FUNCTION_MAX). The number of input samples to reduce across is bounded by MinimumSamplesPerOutput and MaximumSamplesPerOutput.

-field InterpolationMode

Type: DML_INTERPOLATION_MODE

The interpolation mode to use when resizing the regions.

  • DML_INTERPOLATION_MODE_NEAREST_NEIGHBOR. Uses the Nearest Neighbor algorithm, which chooses the input element nearest to the corresponding pixel center for each output element.
  • DML_INTERPOLATION_MODE_LINEAR. Uses the Bilinear algorithm, which computes the output element by doing the weighted average of the 2 nearest neighboring input elements per dimension. Since only 2 dimensions are resized, the weighted average is computed on a total of 4 input elements for each output element.

-field SpatialScaleX

Type: FLOAT

The X (or width) component of the scaling factor to multiply the ROITensor coordinates by in order to make them proportionate to InputHeight and InputWidth. For example, if ROITensor contains normalized coordinates (values in the range [0..1]), then SpatialScaleX would usually have the same value as InputWidth.

-field SpatialScaleY

Type: FLOAT

The Y (or height) component of the scaling factor to multiply the ROITensor coordinates by in order to make them proportionate to InputHeight and InputWidth. For example, if ROITensor contains normalized coordinates (values in the range [0..1]), SpatialScaleY would usually have the same value as InputHeight.

-field OutOfBoundsInputValue

Type: FLOAT

The value to read from InputTensor when the ROIs are outside the bounds of InputTensor. This can happen when the values obtained after scaling ROITensor by SpatialScaleX and SpatialScaleY are bigger than InputWidth and InputHeight.

-field MinimumSamplesPerOutput

Type: UINT

The minimum number of input samples to use for every output element. The operator will calculate the number of input samples by doing ScaledCropSize / OutputSize, and then clamp it to MinimumSamplesPerOutput and MaximumSamplesPerOutput.

-field MaximumSamplesPerOutput

Type: UINT

The maximum number of input samples to use for every output element. The operator will calculate the number of input samples by doing ScaledCropSize / OutputSize, and then clamp it to MinimumSamplesPerOutput and MaximumSamplesPerOutput.

Availability

This operator was introduced in DML_FEATURE_LEVEL_3_0.

Tensor constraints

InputTensor, OutputTensor, and ROITensor must have the same DataType.

Tensor support

DML_FEATURE_LEVEL_5_0 and above

Tensor Kind Supported dimension counts Supported data types
InputTensor Input 4 FLOAT32, FLOAT16
ROITensor Input 2 to 4 FLOAT32, FLOAT16
BatchIndicesTensor Input 1 to 4 UINT64, UINT32
OutputTensor Output 4 FLOAT32, FLOAT16

DML_FEATURE_LEVEL_3_0 and above

Tensor Kind Supported dimension counts Supported data types
InputTensor Input 4 FLOAT32, FLOAT16
ROITensor Input 2 to 4 FLOAT32, FLOAT16
BatchIndicesTensor Input 1 to 4 UINT32
OutputTensor Output 4 FLOAT32, FLOAT16