A replacement for cv2.videoWriter with CUDA support in Linux, which is currently not supported Instead, we use ffmpeg built with CUDA support to encode and save frames.

Tested requirements

• NVIDIA GPU
• CUDA 11.3.r11.3
• Ubuntu 20.04.4 LTS
• Python 3.8.10

Quick encoding test from a random image with 2000x2000 image at 45 FPS

cv2cuda --device gpu --width 2000 --height 2000 --fps 45 --output output.mp4

Implementation details

FFMPEG

Following the guides from NVIDIA, ffmpeg and StackOverflow:

• https://docs.nvidia.com/video-technologies/video-codec-sdk/ffmpeg-with-nvidia-gpu/#commonly-faced-issues-and-tips-to-resolve-them
• https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu
• https://stackoverflow.com/a/55747785/3541756

an ffmpeg subprocess is used to receive Python's standard output and encode it into a video using GPU hardware acceleration. The ffmpeg command has the following structure

ffmpeg -y -r FRAMERATE -f rawvideo -pix_fmt gray -vsync 0 -extra_hw_frames 2 -s WIDTHxHEIGHT -i - -an -c:v h264_nvenc OUTPUT.mp4

• -y -r FRAMERATE tells ffmpeg to ignore and overwrite any existing data in the output and sets the framerate

• -f rawvideo -pix_fmt gray tells ffmpeg that the input will be raw video (which matches what Python outputs), and the pixel format is gray (only gray is supported for now by cv2cuda)

• -vsync 0 -extra_hw_frames 2 are flags that I have read can improve the performance. But I am not sure why and maybe they dont. They could potentially be removed

• -s WIDTHxHEIGHT tells ffmpeg what width and height to expect in the incoming frames

• -i - tells ffmpeg to read input from the standard input (i.e. to listen to Python)

• -an tells ffmpeg there is no audio input

• c:v h264_nvenc tells ffmpeg to use hardware acceleration by using the NVIDIA h264_nvenc codec for encoding

• OUTPUT.mp4 (or whatever path) is the output file

This command is built and run for you when you use cv2cuda

VideoWriter

The cv2cuda package provides a drop-in replacement for the popular cv2.VideoWriter class called cv2cuda.VideoWriter. It has functionality very similar to the original cv2.VideoWriter but uses ffmpeg behind the scenes to make use of the GPU. Thus it is a solution to the issue posed here opencv/opencv_contrib#3044

You can initialize like so:

video_writer = cv2cuda.VideoWriter(
                filename = output_mp4,
                apiPreference="FFMPEG",
                fourcc="h264_nvenc",
                fps=fps,
                frameSize=frameSize,
                isColor=False,
)

i.e. just like the cv2.VideoWriter

You can add frames to the video, and stop it, in a way identical to the cv2.VideoWriter

video_writer.write(frame)
video_writer.release()

Versions

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Sun_Mar_21_19:15:46_PDT_2021
Cuda compilation tools, release 11.3, V11.3.58
Build cuda_11.3.r11.3/compiler.29745058_0

ffmpeg version n4.2.2 Copyright (c) 2000-2019 the FFmpeg developers
built with gcc 9 (Ubuntu 9.3.0-17ubuntu1~20.04)
configuration: --prefix=/usr/local/ffmpeg --pkg-config-flags=--static --enable-nonfree --enable-gpl --enable-version3 --enable-libmp3lame --enable-libvpx --enable-libopus --enable-opencl --enable-libxcb --enable-opengl --enable-nvenc --enable-vaapi --enable-vdpau --enable-ffplay --enable-ffprobe --enable-libxvid --enable-libx264 --enable-libx265 --enable-openal --enable-openssl --enable-cuda-nvcc --enable-cuvid --extra-cflags=-I/usr/local/cuda/include --extra-ldflags=-L/usr/local/cuda/lib64 --nvccflags='-gencode arch=compute_52,code=sm_52 -O2' --enable-pic
libavutil      56. 31.100 / 56. 31.100
libavcodec     58. 54.100 / 58. 54.100
libavformat    58. 29.100 / 58. 29.100
libavdevice    58.  8.100 / 58.  8.100
libavfilter     7. 57.100 /  7. 57.100
libswscale      5.  5.100 /  5.  5.100
libswresample   3.  5.100 /  3.  5.100
libpostproc    55.  5.100 / 55.  5.100