SVO to mp4 export speed up on Nvidia Jetsons


I am recording videos (lossy H265) on Jetson Orin NX 16GB

using the following code

After that I need to perform extraction to an .avi or .mp4 file while applying rotation and blurring of the faces (anonymization) using built-in body tracking (head position). Unfortunately on Jetson Orin NX it takes enormous amount of time (up to x10-15 of the original video duration).

Based on a basic profiling most of the time takes actual encoding. In the official ZED SDK sample cv2.VideoWriter_fourcc is used. I tried to use ffmpeg and Python pipes to speed up the process but it didn’t end with any significant performance boost.

What are the most optimal params for ffmpeg to use Jetson hardware acceleration for video extraction from .svo files ? Is there an alternative way of doing that ?

works but runs entirely on CPU which is very slow

ffmpeg_process = sp.Popen(shlex.split(
    f'ffmpeg -y -s {rotated_width}x{rotated_height} -pixel_format bgr24 -f rawvideo -r {svo_fps} -i pipe: -vcodec libx265 -pix_fmt yuv420p -crf 26 {deided_mp4_filename}'),

doesn’t work

# ffmpeg_process = sp.Popen(shlex.split(
#     f'ffmpeg -hwaccel cuda -y -s {rotated_width}x{rotated_height} -pixel_format bgr24 -f rawvideo -r {svo_fps} -i pipe: -vcodec hevc_nvenc -pix_fmt yuv420p -crf 26 {deided_mp4_filename}'),
#     stdin=sp.PIPE)

Thank you

Hi @efan4ik,

Welcome to the Stereolabs forums :slight_smile:

If performance is what you are looking for, have you tried the export sample in C++ and noticed any differences for exporting the videos?

I am not familiar with ffmpeg options in general, but the workflow you are performing is correct, decoding the svo and re-encoding it again is the way to go.