[FFmpeg-user] Encoding multiple files for adaptive bit rate streaming using RTMP and HLS

Post by Joel Lopez
Can I add a few lower versions to the mix and a higher one?

Yes, you should do that.

Post by Joel Lopez
Can I realign the keyframes of the existing files?

That's going to be hard. You're better off re-encoding from scratch
with a fixed interval.

Post by Joel Lopez
I'm not so sure how crucial aligned key frames are. What have you seen out there?

The recommendation is definitely to keep them aligned; otherwise
you'll run into sync issues.

Post by Joel Lopez
What bitrates and dimensions do you think they're using? What do you
recommend is good for adaptive bit rate switching?

See https://developer.apple.com/library/ios/technotes/tn2224/_index.html#//apple_ref/doc/uid/DTS40009745-CH1-SETTINGSFILES

It's a good starting point.

Post by Joel Lopez
He said I didn't need to change the dimensions when making
different bit rate versions. What's your opinion?

Typically you should vary both bit rate and resolution, and frame rate
if you're targeting really crappy devices.
Changing just the bit rate but not the resolution may make the video
look worse at lower bit rates than if it had a lower resolution and
the same low bit rate. I recently did a test with bitrate-only
switching and I'm not sure if I'd recommend that.

Post by Joel Lopez
How do I change this command to make the different bitrate or sizes?
ffmpeg [rawvideo demuxer options if required] -i input -c:v libx264
-crf 23 -preset medium -pix_fmt yuv420p -c:a libfdk_aac -vbr 5
-movflags +faststart output.mp4

Check the H.264 encoding guide
(https://trac.ffmpeg.org/wiki/Encode/H.264). Do not use the CRF option
but use "-b:v 1M" for 1 MBit/s, or similar. Ideally, do a two-pass
encode for better quality, if you have the time.

Henk D. Schoneveld

2015-05-12 09:47:44 UTC

Verstuurd vanaf mijn iPhone

Post by Joel Lopez
Can I add a few lower versions to the mix and a higher one?

Yes, you should do that.

Post by Joel Lopez
Can I realign the keyframes of the existing files?

That's going to be hard. You're better off re-encoding from scratch
with a fixed interval.

Post by Joel Lopez
I'm not so sure how crucial aligned key frames are. What have you seen out there?

The recommendation is definitely to keep them aligned; otherwise
you'll run into sync issues.

Post by Joel Lopez
What bitrates and dimensions do you think they're using? What do you
recommend is good for adaptive bit rate switching?

See https://developer.apple.com/library/ios/technotes/tn2224/_index.html#//apple_ref/doc/uid/DTS40009745-CH1-SETTINGSFILES
It's a good starting point.

Post by Joel Lopez
He said I didn't need to change the dimensions when making
different bit rate versions. What's your opinion?

Would you be so kind to explain why to NOT use the crf option?

Post by Werner Robitza
_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

Werner Robitza

2015-05-12 11:50:22 UTC

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes. For
streaming purposes, this is not ideal, since your adaptive streaming
client assumes that a segment encoded at a target bitrate of x kBit/s
is can actually be transmitted over a link with x kBit/s throughput.

If you want to make sure you're not exceeding a certain bandwidth, the
VBV encoding mode is probably the best option (see also
https://trac.ffmpeg.org/wiki/EncodingForStreamingSites). Therefore,
set -maxrate and -bufsize.

A single-pass constant bitrate will not generally be more stable than
CRF, but it should be less spikey.

See this chart for a comparison between CRF, single-pass CBR and
single-pass CBR with -maxrate set: Loading Image...

y-axis is the frame size, moving average of 120 frames.

Henk D. Schoneveld

2015-05-12 13:16:21 UTC

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes.

Your goal is max quality within a given link-capacity I assume.
Upfront choosing an arbitrary bitrate to achieve max possible quality seems sub-optimal/contradictionary to me.
A. to many bits for talking heads
B. to few bits for action dominant events.

Post by Werner Robitza
For
streaming purposes, this is not ideal, since your adaptive streaming
client assumes that a segment encoded at a target bitrate of x kBit/s
is can actually be transmitted over a link with x kBit/s throughput.

Where both the stream and link bitrate are averages. A stream consists of I and P and sometimes B frames, the size of these individual frames differ in size. For example choosing a small GOP-size, with relatively more I-frames, will result in more or less 'avoidable' relatively lower average quality.

Post by Werner Robitza
If you want to make sure you're not exceeding a certain bandwidth, the
VBV encoding mode is probably the best option (see also
https://trac.ffmpeg.org/wiki/EncodingForStreamingSites). Therefore,
set -maxrate and -bufsize.
A single-pass constant bitrate will not generally be more stable than
CRF, but it should be less spikey.
See this chart for a comparison between CRF, single-pass CBR and
single-pass CBR with -maxrate set: http://i.imgur.com/GxTW4Jy.png
y-axis is the frame size, moving average of 120 frames.

I see the difference between the methods, but I don’t really understand what it’s trying to tel me. What does the X-axis say, total stream-size/#frames/ ?

Post by Werner Robitza
_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

Werner Robitza

2015-05-12 13:50:43 UTC

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes.

Yes, but that's still what's typically done, unless you choose a CQ
encoding type with a max-rate (which is what libvpx recommends too).

You should've mentioned that you're experienced with this -- otherwise
I would've given a different answer.

Post by Henk D. Schoneveld
I see the difference between the methods, but I don’t really understand what it’s trying to tel me. What does the X-axis say, total stream-size/#frames/ ?

It's the frame number (index) in a 3 minute sequence.

Henk D. Schoneveld

2015-05-12 14:32:17 UTC

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes.

Yes, but that's still what's typically done, unless you choose a CQ
encoding type with a max-rate (which is what libvpx recommends too).

But libvpx is much less efficient than libx264 for the same quality.

Post by Werner Robitza
You should've mentioned that you're experienced with this -- otherwise
I would've given a different answer.

Hm what difference does it make if I know a little bit more or less to how/what you’re answering ?
Another option to ‘optimise’ quality for a given bitrate is to use anamorphic encoding. I know YouTube doesn’t accept that but all modern players I know of handle it without any problem.
For 720p I use -s 880x720 which works for me pretty well. With this you’ll get a reduction of (1280-880)/1280=31% in needed bits for very reasonable quality. At least a lot better then reducing the bitrate by that % and using 1:1 ie. -s 1280x720 encoding.

Post by Henk D. Schoneveld
I see the difference between the methods, but I don’t really understand what it’s trying to tel me. What does the X-axis say, total stream-size/#frames/ ?

It's the frame number (index) in a 3 minute sequence.
_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

Henk D. Schoneveld

2015-05-12 17:36:13 UTC

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes.

Yes, but that's still what's typically done, unless you choose a CQ
encoding type with a max-rate (which is what libvpx recommends too).

But libvpx is much less efficient than libx264 for the same quality.

Post by Werner Robitza
You should've mentioned that you're experienced with this -- otherwise
I would've given a different answer.

Hm what difference does it make if I know a little bit more or less to how/what youre answering ?
Another option to optimise quality for a given bitrate is to use anamorphic encoding. I know YouTube doesnt accept that but all modern players I know of handle it without any problem.
For 720p I use -s 880x720 which works for me pretty well. With this youll get a reduction of (1280-880)/1280=31% in needed bits for very reasonable quality. At least a lot better then reducing the bitrate by that % and using 1:1 ie. -s 1280x720 encoding.

Test results by doing as forementioned
2442734 May 12 2015 Source_Code-crf22-1280.mp4
1656672 May 12 2015 Source_Code-crf22-880.mp4
2045725 May 12 2015 Source_Code-crf23-1280.mp4
1403091 May 12 2015 Source_Code-crf23-880.mp4

This are video only files.

Post by Henk D. Schoneveld
I see the difference between the methods, but I dont really understand what its trying to tel me. What does the X-axis say, total stream-size/#frames/ ?

It's the frame number (index) in a 3 minute sequence.
_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

Joel Lopez

2015-05-12 19:08:02 UTC

Thanks for the advice. Now I guess I should encode a video using
double pass and CFR and compare them. There seems to be a hot debate
about which is best.

The settings below are what the apple.com recommends for 4:3. Forgive
my slowness but I'm having a tough time converting what sites
recommend into ffmpeg commands. Now especially double pass and CFR.

From the wiki page on encoding for streaming they show this example
https://trac.ffmpeg.org/wiki/EncodingForStreamingSites

ffmpeg -i input.mkv -vcodec libx264 -preset medium -maxrate 3000k
-bufsize 6000k \
-vf "scale=1280:-1,format=yuv420p" -g 50 -acodec libmp3lame -b:a 128k
-ac 2 -ar 44100 file.flv

so is this how I should change it? attempting the highest and lowest

ffmpeg -i input.mp4 -vcodec libx264 -preset medium -maxrate 264k -bufsize 500k \
-vf "scale=400:-1,format=yuv420p" -g 50 -acodec aac -b:a 64k -ac 2 -ar
48000 output_400.mp4

ffmpeg -i input.mp4 -vcodec libx264 -preset medium -maxrate 5120k
-bufsize 10000k \
-vf "scale=1280:-1,format=yuv420p" -g 50 -acodec aac -b:a 128k -ac 2
-ar 48000 output_1280.mp4

I'm totally guessing on -bufsize by doubling the -maxrate. What
should it actually be?
Also on the scale, what's the difference between -1 and -2?
Since the -g = 50 does this mean my keyframes will be aligned 5 seconds apart?

400x300
Frame Rate = 12
Total Bit Rate = 264
Video Bit Rate = 200
Audio Bit Rate = 64
Audio Sample Rate = 48
Keyframe = 36
Profile = Baseline, 3.0
B-frames = 0
Segment Size = 9

480x360
Frame Rate = 15
Total Bit Rate = 464
Video Bit Rate = 400
Audio Bit Rate = 64
Audio Sample Rate = 48
Keyframe = 45
Profile = Baseline, 3.0
B-frames = 0
Segment Size = 9

640x480
Frame Rate = 29.97
Total Bit Rate = 664
Video Bit Rate = 600
Audio Bit Rate = 64
Audio Sample Rate = 48
Keyframe = 90
Profile = Baseline, 3.0
B-frames = 0
Segment Size = 9

640x480
Frame Rate = 29.97
Total Bit Rate = 1296
Video Bit Rate = 1200
Audio Bit Rate = 96
Audio Sample Rate = 48
Keyframe = 90
Profile = Baseline, 3.1
B-frames = 0
Segment Size = 9

960x720
Frame Rate = 29.97
Total Bit Rate = 3596
Video Bit Rate = 3500
Audio Bit Rate = 96
Audio Sample Rate = 48
Keyframe = 90
Profile = Main, 3.1
B-frames = as needed
Segment Size = 9

1280x960
Frame Rate = 29.97
Total Bit Rate = 5128
Video Bit Rate = 5000
Audio Bit Rate = 128
Audio Sample Rate = 48
Keyframe = 90
Profile = Main, 3.1
B-frames = as needed
Segment Size = 9

Post by Henk D. Schoneveld
Would you be so kind to explain why to NOT use the crf option?

CRF is essentially a constant quality mode, which results in variable
bitrate depending on the spatiotemporal complexity of the scenes.

Yes, but that's still what's typically done, unless you choose a CQ
encoding type with a max-rate (which is what libvpx recommends too).

But libvpx is much less efficient than libx264 for the same quality.

Post by Werner Robitza
You should've mentioned that you're experienced with this -- otherwise
I would've given a different answer.

Post by Henk D. Schoneveld
I see the difference between the methods, but I don’t really understand what it’s trying to tel me. What does the X-axis say, total stream-size/#frames/ ?

It's the frame number (index) in a 3 minute sequence.
_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

_______________________________________________
ffmpeg-user mailing list
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

Joel Lopez

2015-05-12 19:11:50 UTC