[Feature Request]: Add Image rotation or something like cv::warpAffine for ov::preprocess::PrePostProcessor

[personball]

Request Description

In my application, it needs to inference a 1920x1080 image and its 180-degree rotated copy.
After read all doc of ov::preprocess, I didn't found any step like image rotation.
It's better to run this rotation before model input on GPU than use opencv cv::warpAffine on CPU.

Feature Use Case

for now:

cv::Point2f center(1920 / 2.0, 1080 / 2.0);
cv::Mat rotation_matrix = cv::getRotationMatrix2D(center, 180, 1.0);
cv::Mat img_rotated;
cv::warpAffine(img, img_rotated, rotation_matrix, img.size());

something better:

ppp.input().preprocess()
    .convert_element_type(ov::element::f32)
    .scale({ 255.0, 255.0 , 255.0 })
    .rotate(rotation_matrix) // like this
    .warpAffine(rotation_matrix)// or like this, more affine effect
    .convert_color(ov::preprocess::ColorFormat::RGB);

Issue submission checklist

• The feature request or improvement must be related to OpenVINO

[ilya-lavrenov]

It can be implemented via GridSample operation (see https://docs.openvino.ai/2024/documentation/openvino-ir-format/operation-sets/operation-specs/image/grid-sample-9.html)

#include <openvino/openvino.hpp>

using namespace ov;

// Step 1: Input Node
auto input_shape = ov::Shape{1, 3, 224, 224}; // Batch size, channels, height, width
auto input = std::make_shared<op::v0::Parameter>(element::f32, input_shape);

// Step 2: Affine Transformation Matrix
std::vector<float> affine_matrix_data = {1.0, 0.0, 0.0, 0.0, 1.0, 0.0}; // Identity affine matrix
auto affine_matrix = op::v0::Constant::create(element::f32, {2, 3}, affine_matrix_data);

// Step 3: Convert to 4x4 Transformation Matrix
std::vector<float> matrix_4x4_data = {
    affine_matrix_data[0], affine_matrix_data[1], 0.0f, affine_matrix_data[2],
    affine_matrix_data[3], affine_matrix_data[4], 0.0f, affine_matrix_data[5],
    0.0f,                0.0f,                1.0f, 0.0f,
    0.0f,                0.0f,                0.0f, 1.0f,
};
auto matrix_4x4 = op::v0::Constant::create(element::f32, {4, 4}, matrix_4x4_data);

// Step 4: Grid Sample Operation
auto grid_sample = std::make_shared<op::v9::GridSample>(
    input, matrix_4x4,
    op::v9::GridSample::Attributes{
        op::v9::GridSample::InterpolationMode::BILINEAR,
        op::v9::GridSample::PaddingMode::ZEROS,
        false // align_corners
    });

// Step 5: Finalize and create the model
auto model = std::make_shared<Model>(NodeVector{grid_sample}, ParameterVector{input});

You can use OpenVINO Preprocessing custom operation https://docs.openvino.ai/2024/openvino-workflow/running-inference/optimize-inference/optimize-preprocessing/preprocessing-api-details.html#custom-operations guide to add warpAffine which is based on OpenVINO operations.

In case of success, could you please try to contribute it to OpenVINO preprocessing API?

[personball]

@ilya-lavrenov thank you very much!

and, one more thing, do we have some post process sample code to impl nms(NonMaxSuppression) for model output?