Realtime/zero-latency video stream: what codec parameters to use?
I am no expert by any means when it comes to streaming and video encoding/decoding, and getting the App to a working point was already a nightmare :/ If H264 should be inappropriate we can switch to any other codec supported by gstreamer
This is most probably not a codec-related issue: delays introduced by a codec occur when using B-frames, which IIRC x264 does not use in zero latency mode.
Other delays in real-time streaming applications are your
- encoding delay (e.g. can your phone encode the data fast enough if not you can try to reduce resolution/quality/bitrate)
- network (which on the LAN should not be an issue)
- jitter/playout buffer
I would suggest looking at the playout buffer, perhaps gstreamer has a way to set the duration? Also, the implementation of the playout buffer plays a big role in how real-time you can go e.g. in older versions of VLC one was able to set the network caching parameter really low e.g. in the order of 100ms. However in current versions of VLC this causes the video to not playback as data "arrives" late. ffmpeg on the other hand is better suited to playback real-time data with low latency. I'm not sure how gstreamer compares.
You can try experimenting with ~100ms and then adjust according to how it performs. This is assuming of course that you can set this parameter in gstreamer.
We do live video streaming from desktop PCs to Raspberry Pis, and we spent an enormous amount of time tweaking both the encoding and decoding portions of our system. Unfortunately most libraries and tools have their out-of-the-box settings geared towards trans-coding or general video playback (not live). We ended up writing our own GStreamer element to do the encoding (using vaapi) and our own Raspberry Pi program to do the decoding (using OMX).
I can offer a few thoughts for you, but nothing specific for the Android decoding scenario, unfortunately.
If you're encoding on a powerful desktop, like i3-i7, make sure you add queues for any significant operation: colorspace conversion, scaling, encoding, etc. So in your pipeline, make sure there is a "queue" between "videoconvert" and "x264enc" so they run on seperate threads.
As Ralf mentioned, you probably want to use only P frames, not B frames, and your x264enc settings likely already do this.
We usually favor dropping frames and showing garbage over using a large jitter buffer. We also adjust the QP (quality of the encode) on the fly to be within our networks means. So I'd suggest adding sync=false to your receiving program. You want to render a frame as soon as you have it. This potentially makes your video less smooth, but if you have a large jitter buffer you're always gonna be delayed. Better to adjust the stream to the network and get rid of the buffer. x264enc has "qp-min" and "qp-max" properties you can try.
Try adjusting the "latency" and "drop-on-latency" properties of your rtpjitterbuffer, or try getting rid of it altogether.
One very nasty thing we discovered is that in the Raspberry Pi decoder it seemed to always have some sort of builtin latency, no matter how live-optimized our stream was. It turns out in the h264 stream there is something called a VUI packet that can be used to tell the decoder what type of stream to expect, and when we supplied this packet the decoder reacted very differently.
bitstream_restriction_flag : 1 motion_vectors_over_pic_boundaries_flag : 1 max_bytes_per_pic_denom : 0 max_bits_per_mb_denom : 0 log2_max_mv_length_horizontal : 10 log2_max_mv_length_vertical : 10 num_reorder_frames : 0 max_dec_frame_buffering : 1 --- this makes a huge difference
For reference: https://www.raspberrypi.org/forums/viewtopic.php?t=41053
So in the above VUI settings I tell the decoder that we'll have a max of one P frame that it needs to buffer. It's crazy how much this helped. Of course we had to also make sure our encoder did only send the one P frame. I'm not sure this is possible to do do with x264enc.
This stuff can get pretty scary. Hopefully someone else has the Android video chops to give you a simpler answer!
EDIT: Regarding queues, I don't parameterize them at all, and in a live streaming situation if your queues fill up you need to scale back (resolution, quality, whatever) anyway. In GStreamer the queue element causes GStreamer to launch a new thread to handle the following portion of the pipeline. You just want to make sure your encode/scaling/colorspace conversion elements work in isolation.
gst-launch-1.0 [GET RAW VIDEO DATA] queue [SCALE] queue [COLORSPACE CONVERT] queue [ENCODE] queue [SEND WHEREVER]
The above will give you five threads.
If you get nothing here, my recommendation is to hit up an Android video API subforum or mailing list to see if anyone else has live video going and if so what tweaks they made to their stream and decoder.
--- Addendum 1-5-18
We've also noticed that some streams can fill up the kernel socket buffer and result in packet drops--particularly on large keyframes. So if you have a larger stream I recommend checking the kernel buffer size using sysctl:
sysctl net.core.rmem_max; sysctl net.core.rmem_default
net.core.rmem_max = 212992
net.core.rmem_default = 212992
Append net.core.rmem_max = whatever in /etc/sysctl.conf on your receiving device, and on udpsrc set buffer-size to this new max value. You can tell if you're still seeing drops by running something like this:
watch -d 'cat /proc/net/snmp | grep Udp: '
...or something like this on your receiving pipeline:
export GST_DEBUG=2,rtpjitterbuffer:5
gst-launch-1.0 udpsrc port=5100 buffer-size=825984 ! application/x-rtp,encoding-name=H264,payload=96 ! rtpjitterbuffer latency=200 ! rtph264depay ! h264parse disable-passthrough=true ! queue ! avdec_h264 output-corrupt=true ! queue ! videoconvert ! ximagesink 2>&1 | grep -i "buffer discon
--- Addendum 1-11-19
If you have the means to navigate the patent situation, the openh264 library from Cisco works very nicely. It's very much tuned for live streaming.
https://github.com/cisco/openh264
https://www.openh264.org/BINARY_LICENSE.txt
There is a GStreamer plugin for it under gst-plugins-bad.