OpenCV image to encode?

[rizwanishaq]

Is there any example where we can convert OpenCV image np.array() to h264 encoding?

[niujiabenbeng]

Hi @rizwanishaq

Here is a simple demo. I will be glad if this script could provide some help.

#! /usr/bin/env python
# coding: utf-8

import cv2
import numpy as np
import PyNvCodec as nvc

frame_width  = 640
frame_height = 480
frame_fps    = 30
frame_count  = 300

def cvtColor_BGR2YUV_NV12(image):
    height, width = image.shape[:2]
    yuv = cv2.cvtColor(image, cv2.COLOR_BGR2YUV_I420)
    uuvv = yuv[height:].reshape(2, -1)
    uvuv = np.transpose(uuvv, axes=(1, 0))
    yuv[height:] = uvuv.reshape(-1, width)
    return yuv

# create an encoder
encoder = nvc.PyNvEncoder({
    "preset": "P5",
    "tuning_info": "high_quality",
    "profile": "high",
    "codec": "h264",
    "s": f"{frame_width}x{frame_height}",
    "fps": str(frame_fps),
    "gop": str(frame_fps),
}, gpu_id=0)

with open("output.h264", "wb") as dstfile:
    for i in range(frame_count):
        # generate bgr frame
        bgr = np.full((frame_height, frame_width, 3), i % 256, dtype=np.uint8)
        # convert bgr to nv12
        nv12 = cvtColor_BGR2YUV_NV12(bgr)
        # encode frame
        packet = np.ndarray(shape=(0), dtype=np.uint8)
        if encoder.EncodeSingleFrame(nv12, packet):
            dstfile.write(bytearray(packet))
    # encoder is asynchronous, so we need to flush it
    packet = np.ndarray(shape=(0), dtype=np.uint8)
    while encoder.FlushSinglePacket(packet):
        dstfile.write(bytearray(packet))

[rizwanishaq]

Thanks, working.

[rizwanishaq]

Is there any way that we make this part

def cvtColor_BGR2YUV_NV12(image):
    height, width = image.shape[:2]
    yuv = cv2.cvtColor(image, cv2.COLOR_BGR2YUV_I420)
    uuvv = yuv[height:].reshape(2, -1)
    uvuv = np.transpose(uuvv, axes=(1, 0))
    yuv[height:] = uvuv.reshape(-1, width)
    return yuv

on GPU, as this is using all the CPU.

[niujiabenbeng]

maybe this could help.

#! /usr/bin/env python
# coding: utf-8

import cv2
import numpy as np
import PyNvCodec as nvc

frame_width  = 640
frame_height = 480
frame_fps    = 30
frame_count  = 300

class ColorConverter:

    def __init__(self, width, height, gpuid):
        # vpf does not support rgb_planar -> nv12, so we need the following chain:
        #     rgb_planar -> rgb -> yuv420 -> nv12
        self.to_rgb = nvc.PySurfaceConverter(width, height, nvc.PixelFormat.RGB_PLANAR, nvc.PixelFormat.RGB, gpuid)
        self.to_yuv420 = nvc.PySurfaceConverter(width, height, nvc.PixelFormat.RGB, nvc.PixelFormat.YUV420, gpuid)
        self.to_nv12 = nvc.PySurfaceConverter(width, height, nvc.PixelFormat.YUV420, nvc.PixelFormat.NV12, gpuid)
        self.context = nvc.ColorspaceConversionContext(nvc.ColorSpace.BT_601, nvc.ColorRange.MPEG)
        self.uploader = nvc.PyFrameUploader(width, height, nvc.PixelFormat.RGB_PLANAR, gpuid)

    def convert(self, image):
        # convert opencv-style image to rgb_planar
        image = image[:, :, ::-1].transpose((2, 0, 1))
        image = np.ascontiguousarray(image)
        # copy data: cpu -> gpu
        surface = self.uploader.UploadSingleFrame(image)
        if surface.Empty(): return None
        # do actual convertions
        surface = self.to_rgb.Execute(surface, self.context)
        if surface.Empty(): return None
        surface = self.to_yuv420.Execute(surface, self.context)
        if surface.Empty(): return None
        surface = self.to_nv12.Execute(surface, self.context)
        if surface.Empty(): return None
        return surface

encoder = nvc.PyNvEncoder({
    "preset": "P5",
    "tuning_info": "high_quality",
    "profile": "high",
    "codec": "h264",
    "s": f"{frame_width}x{frame_height}",
    "fps": str(frame_fps),
    "gop": str(frame_fps),
}, gpu_id=0)

converter = ColorConverter(frame_width, frame_height, 0)
with open("output.h264", "wb") as dstfile:
    for i in range(frame_count):
        # generate bgr frame
        bgr = np.full((frame_height, frame_width, 3), i % 256, dtype=np.uint8)
        # convert bgr(cpu) to nv12(gpu)
        nv12 = converter.convert(bgr)
        # encode frame, using: EncodeSingleSurface
        packet = np.ndarray(shape=(0), dtype=np.uint8)
        if encoder.EncodeSingleSurface(nv12, packet):
            dstfile.write(packet.data)
    # encoder is asynchronous, so we need to flush it
    packet = np.ndarray(shape=(0), dtype=np.uint8)
    while encoder.FlushSinglePacket(packet):
        dstfile.write(packet.data)

[rizwanishaq]

yes, it is now working on GPU, thanks.

[rizwanishaq]

"Is it possible to generate H.264 encoded frames, feed them back into a decoder, and then convert them back to cv2 images for testing purposes?"

[RomanArzumanyan]

Hi @rizwanishaq

Please check out https://github.com/RomanArzumanyan/VALI. It’s a VPF successor which is actively developed and maintained. It has compatible API and module naming.

WRT you question - yes, it’s possible. Please check out SampleMeasureVideoQuality in VPF repository or VALI tests.