Quality of video over webRTC with SRS compared to OvenMediaEngine (GPU enabled) about srs

[suzp1984]

Could you gives the video source to test? and the steps to reproduce this issue?

[antonymarion]

sure

forest.mp4

then

fmpeg -d -stream_loop -1 -re -i ./forest.mp4 -c copy -f flv -y rtmp:/{your_ip}/live/test </dev/null

[suzp1984]

ffprobe -show_frames forest.mp4 | grep -i 'pict_type'

There are B frames in this video, while webrtc don't support B frame, which will be drop by default at SRS.
That's the reason of bad quality.

How to workaround

ffmpeg -d -stream_loop -1 -re -i ./forest.mp4 -vcodec h264 -profile:v baseline -f flv -y rtmp:/{your_ip}/live/test </dev/null
or
ffmpeg -d -stream_loop -1 -re -i ./forest.mp4 -vcodec h264 -profile:v high -bf 0 -f flv -y rtmp:/{your_ip}/live/test </dev/null

Both of above publish cmd will resolve your problem.

About the GPU Acceleration question

No, SRS will not do any decode or encode, but just demux the video format, get the nalu elements inside the video frames, for the RTMP which is flv tags, then remux the nalu to the RTP packet. In short, the most common job of a live stream server is just do video format demux -> remux, or just re-streaming.

For the OvenMediaEngine, I didn't read it's source code yet, but he process video stream in this way: demux -> decode -> encode -> remux. Which means he did video decode->encode which is cpu | gpu intense job, and this kinds of server will be quite different than general live media server.

[antonymarion]

Well thks for the clarification.

That was not really intensive on a RTX Titan GPU though (around 1% usage if I remember well..)

I will go back to SRS if I see that OvenMediaEngine is consuming too much resources (not the case right now 😉)