How to undistort I420 image data? Efficiently - opencv

Question

I am able to undistort RGB image successfully.
Now, I am working on directly undistort I420 data, instead of first converting it to RGB.
Below are the steps I followed after camera calibration.
K = cv::Matx33d(541.2152931632737, 0.0, 661.7479652584254,
0.0, 541.0606969363056, 317.4524205037745,
0.0, 0.0, 1.0);
D = cv::Vec4d(-0.042166406281296365, -0.001223961942208027, -0.0017036710622692108, 0.00023929900459453295);
newSize = cv::Size(3400, 1940);
cv::Matx33d new_K;
cv::fisheye::estimateNewCameraMatrixForUndistortRectify(K, D, cv::Size(W, H), cv::Mat::eye(3, 3, CV_64F), new_K, 1, newSize); // W,H are the distorted image size
cv::fisheye::initUndistortRectifyMap(K, D, cv::Mat::eye(3, 3, CV_64F), new_K, newSize, CV_16SC2, mapx, mapy);
cv::remap(src, dst, mapx, mapy, cv::INTER_LINEAR);
Above code is giving me undistorted image successfully.
Now I want to undistort I420 data. So, now my src will be an I420/YV12 data.
How can I undistort an I420 data, without converting it first to RGB?
By the way
I420 is an image format with only 1 channel(unlike 3 channels in RGB). It has height = 1.5*image height. Its width is equal to image width.
Below code is to convert I420 to BGR
cvtColor(src, BGR, CV_YUV2BGR_I420, 3);
BGR - pixel arrangement
I420 - pixel arrangement

Answer 1

The most efficient solution is resizing mapx and mapy and applying shrunk maps on down-sampled U and V channels:
Shrink mapx and mapy by a factor of x2 in each axis - create smaller maps matrices.
Divide all elements of shrank maps by 2 (applies mapping lower resolution image).
Apply mapx and mapy on Y color channel.
Apply shrunk_mapx and shrunk_mapy on down-sampled U and V color channels.
Here is a Python OpenCV sample code (please read the comments):
import cv2 as cv
import numpy as np
# For the example, read Y, U and V as separate images.
srcY = cv.imread('DistortedChessBoardY.png', cv.IMREAD_GRAYSCALE) # Y color channel (1280x720)
srcU = cv.imread('DistortedChessBoardU.png', cv.IMREAD_GRAYSCALE) # U color channel (640x360)
srcV = cv.imread('DistortedChessBoardV.png', cv.IMREAD_GRAYSCALE) # V color channel (640x360)
H, W = srcY.shape[0], srcY.shape[1]
K = np.array([[541.2152931632737, 0.0, 661.7479652584254],
[0.0, 541.0606969363056, 317.4524205037745],
[0.0, 0.0, 1.0]])
D = np.array([-0.042166406281296365, -0.001223961942208027, -0.0017036710622692108, 0.00023929900459453295])
# newSize = cv::Size(3400, 1940);
newSize = (850, 480)
# cv::Matx33d new_K;
new_K = np.eye(3)
# cv::fisheye::estimateNewCameraMatrixForUndistortRectify(K, D, cv::Size(W, H), cv::Mat::eye(3, 3, CV_64F), new_K, 1, newSize); // W,H are the distorted image size
new_K = cv.fisheye.estimateNewCameraMatrixForUndistortRectify(K, D, (W, H), np.eye(3), new_K, 1, newSize)
# cv::fisheye::initUndistortRectifyMap(K, D, cv::Mat::eye(3, 3, CV_64F), new_K, newSize, CV_16SC2, mapx, mapy);
mapx, mapy = cv.fisheye.initUndistortRectifyMap(K, D, np.eye(3), new_K, newSize, cv.CV_16SC2);
# cv::remap(src, dst, mapx, mapy, cv::INTER_LINEAR);
dstY = cv.remap(srcY, mapx, mapy, cv.INTER_LINEAR)
# Resize mapx and mapy by a factor of x2 in each axis, and divide each element in the map by 2
shrank_mapSize = (mapx.shape[1]//2, mapx.shape[0]//2)
shrunk_mapx = cv.resize(mapx, shrank_mapSize, interpolation = cv.INTER_LINEAR) // 2
shrunk_mapy = cv.resize(mapy, shrank_mapSize, interpolation = cv.INTER_LINEAR) // 2
# Remap U and V using shunk maps
dstU = cv.remap(srcU, shrunk_mapx, shrunk_mapy, cv.INTER_LINEAR, borderValue=128)
dstV = cv.remap(srcV, shrunk_mapx, shrunk_mapy, cv.INTER_LINEAR, borderValue=128)
cv.imshow('dstY', dstY)
cv.imshow('dstU', dstU)
cv.imshow('dstV', dstV)
cv.waitKey(0)
cv.destroyAllWindows()
Result:
Y:
U:
V:
After converting to RGB:
C++ implementation considerations:
Since I420 format arranges Y, U and V as 3 continuous planes in memory, it's simple to set a pointer to each "plane", and treat it as a Grayscale image.
Same data ordering applies the output image - set 3 pointer to output "planes".
Illustration (assuming even width and height, and assume byte stride equals width):
srcY -> YYYYYYYY dstY -> YYYYYYYYYYYY
YYYYYYYY YYYYYYYYYYYY
YYYYYYYY YYYYYYYYYYYY
YYYYYYYY YYYYYYYYYYYY
YYYYYYYY remap YYYYYYYYYYYY
YYYYYYYY ======> YYYYYYYYYYYY
srcU -> UUUU YYYYYYYYYYYY
UUUU dstU -> YYYYYYYYYYYY
UUUU UUUUUU
srcV -> VVVV UUUUUU
VVVV UUUUUU
VVVV UUUUUU
dstV -> VVVVVV
VVVVVV
VVVVVV
VVVVVV
Implementing above illustration is C++
Under the assumption that width and height are even, and byte stride equals width, you can use the following C++ example for converting I420 to Y, U and V planes:
Assume: srcI420 is Wx(H*3/2) matrix in I420 format, like cv::Mat srcI420(cv::Size(W, H * 3 / 2), CV_8UC1);.
int W = 1280, H = 720; //Assume resolution of Y plane is 1280x720
//Pointer to Y plane
unsigned char *pY = (unsigned char*)srcI420.data;
//Y plane as cv::Mat, resolution of srcY is 1280x720
cv::Mat srcY = cv::Mat(cv::Size(W, H), CV_8UC1, (void*)pY);
//U plane as cv::Mat, resolution of srcU is 640x360 (in memory buffer, U plane is placed after Y).
cv::Mat srcU = cv::Mat(cv::Size(W/2, H/2), CV_8UC1, (void*)(pY + W*H));
//V plane as cv::Mat, resolution of srcV is 640x360 (in memory buffer, V plane is placed after U).
cv::Mat srcV = cv::Mat(cv::Size(W / 2, H / 2), CV_8UC1, (void*)(pY + W*H + (W/2*H/2)));
//Display srcY, srcU, srcV for testing
cv::imshow("srcY", srcY);
cv::imshow("srcU", srcU);
cv::imshow("srcV", srcV);
cv::waitKey(0);
Above example uses pointer manipulations, without the need for copying the data.
You can use the same pointer manipulations for your destination I420 image.
Note: The solution is going to work in most cases, but not guaranteed to work in all cases.

Answer 2

EDIT: Components are not interleaved in the YV12 format, so the following will not work:
If the YV12 data is a one channel image, the interpolation of the remap operation is applied to the value represented by all three YUV data instead of individual Y, U and V components.
Therefore, roughly speaking, instead of doing a
c.YYYYYYYY, c.UU, c.VV
it will perform a
c.YYYYYYYYUUVV
during a linear interpolation.
You can perform a YV12 -> BGR color conversion after remap, but the colors of the interpolated pixels would be wrong.
Instead of doing a linear interpolation, try using a nearest-neighbor interpolation in remap. Then you should be able to get correct colors after YV12 -> BGR color conversion.
So, find mapx, mapy, then remap using INTER_NEAREST, and finally perform a YV12 -> BGR color conversion.