Converting image from YV12 to NV12 with OpenCV

Question

I have a webcam from which I read frames in NV12 format. I convert the frames to RGB, then to YV12, and my goal is to convert them back to NV12, for verification purposes. I am doing something like this:

cv::cvtColor(InputFrame, InputRGB, cv::COLOR_YUV2RGB_NV12);
cv::cvtColor(InputRGB, OutputYV12, cv::COLOR_RGB2YUV_YV12);

I wrote the following function to convert from YV12 to NV12 (similar to this post - Convert YV12 to NV21 (YUV YCrCb 4:2:0)), which doesn't seem to work. I get a grayscale image with a vague magenta copy mixed over the top half with a vague green copy mixed over the bottom half of my resulting image.

In my function below I'm assuming a layout where the V-plane is sitting next to the U-plane in the matrix. I don't know if that is correct. I first tried following the layout for YV12 as shown at https://learn.microsoft.com/en-us/windows/win32/medfound/recommended-8-bit-yuv-formats-for-video-rendering where the U/V planes sit underneath each other instead of next to each other, but that caused a crash.

void YV12toNV12(const cv::Mat& input, cv::Mat& output, int width, int height) {

        input.copyTo(output);

        for (int row = 0; row < height/2; row++) {
                for (int col = 0; col < width/2; col++) {
                        output.at<uchar>(height + row, 2 * col) = input.at<uchar>(height + row, col);
                        output.at<uchar>(height + row, 2 * col + 1) = input.at<uchar>(height + row, width/2 + col);
                }
        }
}

Any hints appreciated.

Answer 1

Applying the conversion using indexing is confusing.
My suggestion is treating the YV12 image as 3 separate images.

• Y (width x height) - Top Image.
• V (width/2 x height/2) - Below Y
• U (width/2 x height/2) - Below V

---

According to the following documentation:

YV12 is exactly like I420, but the order of the U and V planes is reversed.

I420 is ordered as here:

BGR to I420 conversion is supported by OpenCV, and more documented format compared to YV12, so we better start testing with I420, and continue with YV12 (by switching U and V channels).

---

The main idea is "wrapping" V and U matrices with cv:Mat objects (setting matrix data pointer by adding offsets to the input data pointer).

• inV Comes after Y (half resolution in each axis, and half stride):

    cv::Mat inV = cv::Mat(cv::Size(width/2, height/2), CV_8UC1, (unsigned char*)input.data + stride*height, stride/2);

• inU Comes after V (half resolution in each axis, and half stride):

    cv::Mat inU = cv::Mat(cv::Size(width/2, height/2), CV_8UC1, (unsigned char*)input.data + stride*height + (stride/2)*(height/2), stride/2);

---

Here is the conversion function:

void YV12toNV12(const cv::Mat& input, cv::Mat& output) {
    int width = input.cols;
    int height = input.rows * 2 / 3;
    int stride = (int)input.step[0];    //Rows bytes stride - in most cases equal to width

    input.copyTo(output);

    //Y Channel
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY

    //V Input channel
    // VVVVVVVV
    // VVVVVVVV
    // VVVVVVVV
    cv::Mat inV = cv::Mat(cv::Size(width / 2, height / 2), CV_8UC1, (unsigned char*)input.data + stride * height, stride / 2);   // Input V color channel (in YV12 V is above U).

    //U Input channel
    // UUUUUUUU
    // UUUUUUUU
    // UUUUUUUU
    cv::Mat inU = cv::Mat(cv::Size(width / 2, height / 2), CV_8UC1, (unsigned char*)input.data + stride * height + (stride / 2)*(height / 2), stride / 2);  //Input V color channel (in YV12 U is below V).

    for (int row = 0; row < height / 2; row++) {
        for (int col = 0; col < width / 2; col++) {
            output.at<uchar>(height + row, 2 * col) = inU.at<uchar>(row, col);
            output.at<uchar>(height + row, 2 * col + 1) = inV.at<uchar>(row, col);
        }
    }
}

---

Implementation and Testing:

Creating NV12 sample image using FFmpeg command line tool:

ffmpeg -y -f lavfi -i testsrc=size=192x108:rate=1:duration=1 -pix_fmt nv12 -f rawvideo test.nv12
ffmpeg -y -f rawvideo -pixel_format gray -video_size 192x162 -i test.nv12 -pix_fmt gray test_nv12.png

---

Creating YV12 sample image using MATLAB (or OCTAVE):

NV12 = imread('test_nv12.png');
Y = NV12(1:108, :);
U = NV12(109:end, 1:2:end);
V = NV12(109:end, 2:2:end);

f = fopen('test.yv12', 'w');
fwrite(f, Y', 'uint8');
fwrite(f, V', 'uint8');
fwrite(f, U', 'uint8');
fclose(f);

f = fopen('test.yv12', 'r');
I = fread(f, [192, 108*1.5], '*uint8')';
fclose(f);
imwrite(I, 'test_yv12.png');

---

C++ implementation (both I420toNV12 and YV12toNV12):

#include "opencv2/opencv.hpp"

void YV12toNV12(const cv::Mat& input, cv::Mat& output) {
    int width = input.cols;
    int height = input.rows * 2 / 3;
    int stride = (int)input.step[0];    //Rows bytes stride - in most cases equal to width

    input.copyTo(output);

    //Y Channel
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY

    //V Input channel
    // VVVVVVVV
    // VVVVVVVV
    // VVVVVVVV
    cv::Mat inV = cv::Mat(cv::Size(width / 2, height / 2), CV_8UC1, (unsigned char*)input.data + stride * height, stride / 2);   // Input V color channel (in YV12 V is above U).

    //U Input channel
    // UUUUUUUU
    // UUUUUUUU
    // UUUUUUUU
    cv::Mat inU = cv::Mat(cv::Size(width / 2, height / 2), CV_8UC1, (unsigned char*)input.data + stride * height + (stride / 2)*(height / 2), stride / 2);  //Input V color channel (in YV12 U is below V).

    for (int row = 0; row < height / 2; row++) {
        for (int col = 0; col < width / 2; col++) {
            output.at<uchar>(height + row, 2 * col) = inU.at<uchar>(row, col);
            output.at<uchar>(height + row, 2 * col + 1) = inV.at<uchar>(row, col);
        }
    }
}


void I420toNV12(const cv::Mat& input, cv::Mat& output) {
    int width = input.cols;
    int height = input.rows * 2 / 3;
    int stride = (int)input.step[0];    //Rows bytes stride - in most cases equal to width
    
    input.copyTo(output);

    //Y Channel
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    // YYYYYYYYYYYYYYYY
    
    //U Input channel
    // UUUUUUUU
    // UUUUUUUU
    // UUUUUUUU
    cv::Mat inU = cv::Mat(cv::Size(width / 2, height / 2), CV_8UC1, (unsigned char*)input.data + stride * height, stride / 2);   // Input U color channel (in I420 U is above V).

    //V Input channel
    // VVVVVVVV
    // VVVVVVVV
    // VVVVVVVV
    cv::Mat inV = cv::Mat(cv::Size(width/2, height/2), CV_8UC1, (unsigned char*)input.data + stride*height + (stride/2)*(height/2), stride/2);  //Input V color channel (in I420 V is below U).

    for (int row = 0; row < height / 2; row++) {
        for (int col = 0; col < width / 2; col++) {
            output.at<uchar>(height + row, 2 * col) = inU.at<uchar>(row, col);
            output.at<uchar>(height + row, 2 * col + 1) = inV.at<uchar>(row, col);
        }
    }
}


int main()
{   
    //cv::Mat input = cv::imread("test_I420.png", cv::IMREAD_GRAYSCALE);
    //cv::Mat output;
    //I420toNV12(input, output);
    //cv::imwrite("output_NV12.png", output);

    cv::Mat input = cv::imread("test_YV12.png", cv::IMREAD_GRAYSCALE);
    cv::Mat output;

    YV12toNV12(input, output);
    cv::imwrite("output_NV12.png", output);

    cv::imshow("input", input);
    cv::imshow("output", output);
    cv::waitKey(0);
    cv::destroyAllWindows();
}

---

Testing the output using MATLAB (or OCTAVE):

A = imread('test_nv12.png');
B = imread('output_NV12.png');
display(isequal(A, B))

---

Input (YV12 as grayscale image):

Input (NV12 as grayscale image):