Does opencv or other libraries provide I422(YUV422 Planar) picture format support?

Question

Currently I'm working with some YUV422p-formatted pics. I read the documentations and tried running python codes below:

print([i for i in dir(cv2) if i.startswith("COLOR_YUV")])

And I found no support for YUV422p format. Does anybody know the reason why opencv has no support for such format, or if there is any other python-based libraries that provides such support?

Answer 1

We may use PyAV library.
PyAV is a Pythonic binding for the FFmpeg libraries.

FFmpeg supports YUV422 Planar picture format, and PyAV also supports it.
Unlike other FFmpeg bindings, PyAV is relatively low-level - instead of using FFmpeg CLI, it allows more low level functionalities that supported by the C interface of FFmpeg.

libswscale is "sub module" of FFmpeg that supports multiple color formats conversions.
PyAV exposes some of the capabilities of libswscale.

---

Start by creating raw input file in yuv422p pixel format, using FFmpeg CLI.

Execute the following shell command for building sample input (used for testing):

ffmpeg -y -f lavfi -i testsrc=192x108:rate=1:duration=10 -vf "scale=out_color_matrix=bt709:out_range=tv" -pix_fmt yuv422p -f rawvideo input.yuv422p

---

The next example, uses PyAV for converting from raw yuv422p to BGR pixel format:

• Create a VideoFrame object:

   frame = VideoFrame(cols, rows, 'yuv422p')
  

• Assume y, u and v are 3 color planes (y is full resolution, and u and v are half resolution in the horizontal axis).
  The planes may be stored in NumPy arrays or bytes arrays (multiple buffer formats are supported).
  Update the 3 planes of frame:

   frame.planes[0].update(y)
   frame.planes[1].update(u)
   frame.planes[2].update(v)
  

Convert the frame to BGR format using reformat method, and convert to NumPy array:

    bgr_frame = frame.reformat(cols, rows, 'bgr24', src_colorspace='ITU709', dst_colorspace='DEFAULT', interpolation='BILINEAR')
    bgr = bgr_frame.to_ndarray()

---

The following code sample reads yuv422p raw data from file, converts to BRG using PyAV (and shows the output using OpenCV):

import cv2
from av.video.frame import VideoFrame

# Build raw input file using shell command:  
# ffmpeg -y -f lavfi -i testsrc=192x108:rate=1:duration=10 -vf "scale=out_color_matrix=bt709:out_range=tv" -pix_fmt yuv422p -f rawvideo input.yuv422p

cols, rows, n_frames = 192, 108, 10

frame = VideoFrame(cols, rows, 'yuv422p')

with open('input.yuv422p', 'rb') as f:
    for n in range(n_frames):
        y = f.read(cols*rows)  # y = np.frombuffer(f.read(cols*rows), np.uint8).reshape(rows, cols)
        u = f.read(cols*rows//2)  # u = np.frombuffer(f.read(cols*rows//2), np.uint8).reshape(rows, cols//2)
        v = f.read(cols*rows//2)  # v = np.frombuffer(f.read(cols*rows//2), np.uint8).reshape(rows, cols//2)
        frame.planes[0].update(y)
        frame.planes[1].update(u)
        frame.planes[2].update(v)

        bgr_frame = frame.reformat(cols, rows, 'bgr24', src_colorspace='ITU709', dst_colorspace='DEFAULT', interpolation='BILINEAR')

        bgr = bgr_frame.to_ndarray()

        cv2.imshow('bgr', bgr)
        cv2.waitKey(1000)

cv2.destroyAllWindows()

---

Notes:
Installing PyAV may be problematic (at least when using Windows).
Instead of using pip install, I downloaded a whl file from here (and used pip install for installing the wheel).

I assumed that the YUV applies BT.709 standard (and applies "TV range" - Y range is [16, 235] not [0, 255]).
For correct conversion, you have to know the exacts standard of your raw data.