OpenCV-related Backends

Overview

The max() of the input range for the following backends is all 1.

Name	Initialization Parameters	Input [Type]	Output [Type]	Remarks
DecodeMat	None	data[str/bytes]	result[cv::Mat] color[str]
cvtColorMat	color	data[cv::Mat] color[str]	result[cv::Mat]
ResizeMat	resize_h,resize_w	data[cv::Mat]	result[cv::Mat]
PillowResizeMat	resize_h,resize_w	data[cv::Mat]	result[cv::Mat]	CV_8UC3
ResizePadMat	max_h,max_w,pad_value	data[cv::Mat]	- result[cv::Mat] - inverse_trans [std::function<std::pair<float, float>(float x, float y)>]	CV_8UC3
Mat2Tensor	device=cpu,gpu	data[cv::Mat]	result[at::Tensor]	Mat2CpuTensor Mat2GpuTensor
SaveMat	save_dir	data[cv::Mat]	result[cv::Mat]
AlignMat	align,resize_h,max_w	data[cv::Mat]	result[cv::Mat],resize_h[int],resize_w[int]	max()=UINT32_MAX
DynamicResizeMat		data[cv::Mat],resize_h,resize_w[int]	result[cv::Mat]
PerspectiveTransformMat		data[cv::Mat],TASK_BOX_KEY	result[cv::Mat]

DecodeMat

• Calls OpenCV to decode an image into a cv::Mat using the IMREAD_COLOR mode. The output image will be converted to a 3-channel BGR image.
• If the decoded image is empty, the output will have no result key.
• The color is fixed as "bgr".

cvtColorMat

• The color parameter is the target color space during initialization, and the color space of the input data during processing. If they are different, a color space conversion will be performed, otherwise the input value will be returned.
• The color parameter currently supports "rgb" and "bgr".
• When performing a color space conversion, the input data must be three-channel data, otherwise an exception will be thrown.

ResizeMat

• Calls OpenCV to perform a resize operation using the default interpolation method.
• resize_h and resize_w must be integers and within the valid range of [1, 1024 * 1024].

PillowResizeMat

• The input cv::Mat type must be CV_8UC3.
• resize_h and resize_w must be integers and within the valid range of [1, 1024 * 1024].
• Strictly maintains consistency with the bilinear interpolation results of Pillow (verified on a large amount of data).

ResizePadMat

• Maintains aspect ratio during resize and pads with a constant pad_value aligned to the top left corner.
• The input cv::Mat type must be CV_8UC3.
• max_h and max_w must be integers and within the valid range of [1, 1024 * 1024].
• pad_value supports integers.
• inverse_trans: used to map new coordinates to original coordinates.

Mat2Tensor

• Convert cv::Mat to at::Tensor while keeping the data type unchanged.
• Output shape is 1chw.
• Data is deep copied.
• Similar to SyncTensor, it will synchronize with the current stream.

caution

Please insert stream management operations: Sequential[Mat2Tensor,SyncTensor] or SyncTensor[Mat2Tensor], otherwise Mat2Tensor will use the default cuda stream.

Mat2CpuTensor

• Convert cv::Mat to at::Tensor(cpu) while keeping the data type unchanged.
• Output shape is 1chw.
• Data is deep copied.

Mat2GpuTensor

• Convert cv::Mat to at::Tensor(gpu) while keeping the data type unchanged.
• Output shape is 1chw.

SaveMat

• save_dir: Directory to save the file, which needs to be created in advance.
• The file name suffix is .png, and the name is unique.

AlignMat

Calculate the target length and width of the input image under the following conditions:

• Keep the aspect ratio basically the same.
• The target height is fixed to resize_h.
• The target width is no greater than max_w.
• The length and width are aligned with align.

Initialization

Parameters

• resize_h - Target length.
• max_w - Maximum width.
• align - default=32, the target length and width need to be multiples of this parameter.

Forward

Parameters

• data - Input data, needs to be of type cv::Mat.
• result - Output, result=data.
• resize_h - Output, target height.
• resize_w - Output, target width.

min()/max()

[1, UINT32_MAX]

DynamicResizeMat

Resize the target image according to the resize_h and resize_w provided by the forward input data.

PerspectiveTransformMat

Perform perspective transformation.

Forward

Parameters

• data - Input data, needs to be of type cv::Mat.
• TASK_BOX_KEY - Original coordinates, type is std::vector<std::vector<int>>, external and internal lengths are 4 and 2 respectively, representing four points, and each point has two coordinate values. The order is [top-left, top-right, bottom-right, bottom-left].
• result - Output.