Resize Crop Convert

group NVCV_C_ALGORITHM__RESIZE_CROP

Unnamed Group

NVCVStatus cvcudaResizeCropConvertReformatSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVTensorHandle in, NVCVTensorHandle out, const NVCVSize2D resizeDim, const NVCVInterpolationType interpolation, const int2 cropPos, const NVCVChannelManip manip, const float scale, const float offset, const bool srcCast)

Executes the fused ResizeCropConvertReformat operation on the given cuda stream. This operation does not wait for completion.

ResizeCropConvertReformat is a fused operator that performs the following operations in order:

  1. Resize either a single tensor or each image in an ImageBatchVarShape to a specified width and height (other dimensions are unchanged). This step is identical to the stand-alone Resize operation with the exception of optionally not type-casting the interpolation results back to the input data type (see the srcCast parameter for details).

  2. Crops a specified region out of the resized tensor.

  3. Apply a scale and offset to output result (after resizing and cropping). This can be used to normalize to a new range of values. For example, if the input is unsigned 8-bit values and the output is floating point, setting scale = 1.0/127.5 and offset = -1.0 will convert the 8-bit input values (ranging from 0 to 255) to floating point output values between -1.0 and 1.0.

  4. Optional channel manipulation—i.e., re-order the channels of a tensor (e.g., RGB to BGR). Limited options available.

  5. Convert the element data type to the output tensor’s data type. For example, convert uchar elements to float. Limited options availble.

  6. If output tensor’s layout doesn’t match the input’s layout, reshape the layout to match output layout (e.g., NHWC to NCHW). Limited options available.

NOTES:

  • Since all images in an ImageBatchVarShape are resized to the same size, the resulting collection now fits in a single tensor.

  • Except for nearest-neighbor interpolation (NVCV_INTERP_NEAREST), interpolation (e.g., NVCV_INTERP_LINEAR, NVCV_INTERP_CUBIC, and NVCV_INTERP_AREA) computes resized pixel values using floating point math. However, the stand-alone resize operation (i.e., running the standard Resize operator independently) converts interpolated pixel values back to the source data type since its input and output types must be the same. As an option, this fused operator can either cast the resized pixel values back to the source type (to match results from running the steps independently), or leave them in the interpolated floating-point space to avoid quantization issues that occur from casting back to an integer source type (e.g., uchar). See the srcCast parameter for details.

Limitations:

Input:

  • Data Layout: [NVCV_TENSOR_HWC, NVCV_TENSOR_NHWC]

  • Channels: [1, 3]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

No

16bit Unsigned

No

16bit Signed

No

32bit Unsigned

No

32bit Signed

No

32bit Float

No

64bit Float

No

Output:

  • Data Layout: [NVCV_TENSOR_NHWC, NVCV_TENSOR_HWC, NVCV_TENSOR_NCHW, NVCV_TENSOR_CHW]

  • Channels: [1, 3]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

No

16bit Unsigned

No

16bit Signed

No

32bit Unsigned

No

32bit Signed

No

32bit Float

Yes

64bit Float

No

Input/Output dependency: 

 Property      |  Input == Output
-------------- | -------------
 Data Layout   | No (Limited)
 Data Type     | No (Limited)
 Number        | Yes
 Channels      | Yes
 Width         | No
 Height        | No

Parameters:

  • handle – [in] Handle to the operator. Must not be NULL.

  • stream – [in] Handle to a valid CUDA stream.

  • in – [in] Input tensor or image batch. The images in an image batch can be of different sizes, but all images must have the same data type, channels, and layout.

  • resizeDim – [in] Dimensions, {width, height}, that tensor or image batch images are resized to prior to cropping, see NVCVSize2D for more details.

  • interpolation – [in] Interpolation method to be used, (see NVCVInterpolationType). Currently, only NVCV_INTERP_NEAREST and NVCV_INTERP_LINEAR are available.

  • cropPos – [in] Crop position, (x, y), specifying the top-left corner of the crop region. The crop region’s width and height is specified by the output tensor’s width & height. The crop must fall within the resized image. Let (x, y, w, h) represent the crop rectangle, where x & y are the cropPos coordinates and w and h are the output tensor’s width and height, respectively, then it must be true that:

    • x >= 0,

    • y >= 0,

    • x + w <= resizeDim.w, and

    • y + h <= resizeDim.h.

  • manip – [in] Channel manipulation to be used&#8212;e.g., reshuffle RGB to BGR (see NVCVChannelManip).

  • scale – [in] Scale (i.e., multiply) the resized and cropped output values by this amount. 1.0 results in no scaling of the output values.

  • offset – [in] Offset (i.e., add to) the output values by this amount. This is applied after scaling&#8212;if v is a resized and cropped value, then scale * v + offset is the final output value. 0.0 results in no offset being added to the output.

  • srcCast – [in] Boolean value indicating whether or not the interpolation results during the resize are re-cast back to the input (or source) data type. Most interpolation methods (e.g., NVCV_INTERP_LINEAR) compute resized pixel values using floating point math. This parameter determines if the interpolation result is cast to the source data type before computing the remaining steps in this operator:

    • true: the interpolation result is cast back to the source type prior to computing the remaining steps &#8212; as if calling the stand-alone Resize operator (since its input and output types must be the same). Note: this option can produce quantized outputs (e.g., the input source type is uchar3), even if the destination data type is floating point.

    • false: the interpolation result is NOT cast back to the source type. Rather, the floating-point interpolation results are directly passed on to the remaining steps in the fused operator.

    • Note: in either case (true or false) the final (fused) result is still cast to the destination data type before writing values into the output tensor.

  • out – [out] Output tensor. In addition to the output tensor determining the crop width and height, the output tensor also specifies the data type (e.g., uchar3 or float) and tensor layout (NHWC or NCHW), with limitations.

Return values:

  • NVCV_ERROR_INVALID_ARGUMENT – Some parameter is invalid or outside valid range.

  • NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

  • NVCV_SUCCESS – Operation executed successfully.

NVCVStatus cvcudaResizeCropConvertReformatVarShapeSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVImageBatchHandle in, NVCVTensorHandle out, const NVCVSize2D resizeDim, const NVCVInterpolationType interpolation, const int2 cropPos, const NVCVChannelManip manip, const float scale, const float offset, const bool srcCast)

Functions

NVCVStatus cvcudaResizeCropConvertReformatCreate(NVCVOperatorHandle *handle)

Constructs and an instance of the ResizeCropConvertReformat operator.

Parameters:

handle – [out] Where the image instance handle will be written to.

  • Must not be NULL.

Return values:

  • NVCV_ERROR_INVALID_ARGUMENT – Handle is null.

  • NVCV_ERROR_OUT_OF_MEMORY – Not enough memory to create the operator.

  • NVCV_SUCCESS – Operation executed successfully.