streamer default h265

Alex Kushleyev

@bendraper , unfortunately we have not tried using voxl-streamer for encoding h265. voxl-camera-server does support h265 encoding and by the way we recently fixed the CBR functionality : https://docs.modalai.com/voxl-camera-server/#constant-quantization-vs-constant-bitrate-mode

Do you need to encode an image that you modify outside of the camera server?

If you are willing to share the changes you made while trying to get it to work, i can test it and see if I can spot anything that's easy to fix (you could provide a patch or link to a fork). If we can get it to work, it could be a nice extension of voxl-streamer.

Alex

bendraper

@Alex-Kushleyev Thanks for getting back. Here's our fork of voxl-streamer and the specific commit

As you can see, not much going on here. I just noticed how the 265 omx encoder object would be created but just unused when encoding a raw stream (using hires_small_grey in my tests which is RAW8).

When I try to connect to the stream, streamer gives the following

voxl2-mini:~$ voxl-streamer -i hires_small_grey -s -b 2000000 -p 554
Waiting for pipe hires_small_grey to appear
Found Pipe
detected following stats from pipe:
w: 1024 h: 768 fps: 30 format: RAW8
Stream available at rtsp://127.0.0.1:554/live
A new client rtsp://169.254.4.2:40732(null) has connected, total clients: 1
Camera server Connected
gbm_create_device(156): Info: backend name is: msm_drm
rtsp client disconnected, total clients: 0
no more rtsp clients, closing source pipe intentionally

and using ffmpeg to connect gives the following

ffplay version 4.4.2-0ubuntu0.22.04.1 Copyright (c) 2003-2021 the FFmpeg developers
  built with gcc 11 (Ubuntu 11.2.0-19ubuntu1)
  configuration: --prefix=/usr --extra-version=0ubuntu0.22.04.1 --toolchain=hardened --libdir=/usr/lib/x86_64-linux-gnu --incdir=/usr/include/x86_64-linux-gnu --arch=amd64 --enable-gpl --disable-stripping --enable-gnutls --enable-ladspa --enable-libaom --enable-libass --enable-libbluray --enable-libbs2b --enable-libcaca --enable-libcdio --enable-libcodec2 --enable-libdav1d --enable-libflite --enable-libfontconfig --enable-libfreetype --enable-libfribidi --enable-libgme --enable-libgsm --enable-libjack --enable-libmp3lame --enable-libmysofa --enable-libopenjpeg --enable-libopenmpt --enable-libopus --enable-libpulse --enable-librabbitmq --enable-librubberband --enable-libshine --enable-libsnappy --enable-libsoxr --enable-libspeex --enable-libsrt --enable-libssh --enable-libtheora --enable-libtwolame --enable-libvidstab --enable-libvorbis --enable-libvpx --enable-libwebp --enable-libx265 --enable-libxml2 --enable-libxvid --enable-libzimg --enable-libzmq --enable-libzvbi --enable-lv2 --enable-omx --enable-openal --enable-opencl --enable-opengl --enable-sdl2 --enable-pocketsphinx --enable-librsvg --enable-libmfx --enable-libdc1394 --enable-libdrm --enable-libiec61883 --enable-chromaprint --enable-frei0r --enable-libx264 --enable-shared
  libavutil      56. 70.100 / 56. 70.100
  libavcodec     58.134.100 / 58.134.100
  libavformat    58. 76.100 / 58. 76.100
  libavdevice    58. 13.100 / 58. 13.100
  libavfilter     7.110.100 /  7.110.100
  libswscale      5.  9.100 /  5.  9.100
  libswresample   3.  9.100 /  3.  9.100
  libpostproc    55.  9.100 / 55.  9.100
[rtsp @ 0x74e358000cc0] method DESCRIBE failed: 503 Service Unavailable
rtsp://169.254.4.1:554/live: Server returned 5XX Server Error reply

bendraper

@Alex-Kushleyev it's probably just my gstreamer ignorance and it not being as simple to set up an h265 stream as swapping our h264 with h265

Alex Kushleyev

@bendraper , you got 90% there!

I just tested your version and was able to reproduce the issue, but the following change fixes it:

from

"profile", G_TYPE_STRING, "baseline",

to

"profile", G_TYPE_STRING, "main",

There is no baseline profile in h265. Please try it - i was able to stream h265 with modified voxl-streamer and play using ffplay.

full diff below. We will test more and probably mainline this soon.

Alex

diff --git a/src/pipeline.c b/src/pipeline.c
index 53f47b7..5f88315 100644
--- a/src/pipeline.c
+++ b/src/pipeline.c
@@ -398,9 +398,9 @@ GstElement *create_custom_element(GstRTSPMediaFactory *factory, const GstRTSPUrl
     // g_object_set(context->h264_parser, "alignment", "nal", NULL);
 
     // Configure the OMX encoder
-    g_object_set(context->omx_encoder, "control-rate", 1, 
+    g_object_set(context->omx_h265_encoder, "control-rate", 1, 
                                        "interval-intraframes", 30, NULL);
-    g_object_set(context->omx_encoder, "target-bitrate",
+    g_object_set(context->omx_h265_encoder, "target-bitrate",
                  context->output_stream_bitrate, NULL);
 
     // Configure the RTP input queue
@@ -443,10 +443,10 @@ GstElement *create_custom_element(GstRTSPMediaFactory *factory, const GstRTSPUrl
         gst_caps_unref(filtercaps);
 
         // Configure the caps filter to reflect the output of the OMX encoder
-        filtercaps = gst_caps_new_simple("video/x-h264",
+        filtercaps = gst_caps_new_simple("video/x-h265",
                                         "width", G_TYPE_INT, context->output_stream_width,
                                         "height", G_TYPE_INT, context->output_stream_height,
-                                        "profile", G_TYPE_STRING, "baseline",
+                                        "profile", G_TYPE_STRING, "main",
                                         NULL);
     }
     if ( ! filtercaps) {
@@ -482,10 +482,10 @@ GstElement *create_custom_element(GstRTSPMediaFactory *factory, const GstRTSPUrl
                         context->video_rotate,
                         context->video_rotate_filter,
                         context->encoder_queue,
-                        context->omx_encoder,
+                        context->omx_h265_encoder,
                         context->rtp_filter,
                         context->rtp_queue,
-                        context->rtp_payload,
+                        context->rtp_h265_payload,
                         NULL);
     }
 
@@ -536,10 +536,10 @@ GstElement *create_custom_element(GstRTSPMediaFactory *factory, const GstRTSPUrl
         last_element = context->video_rotate_filter;
         if ( ! gst_element_link_many(last_element,
                                      context->encoder_queue,
-                                     context->omx_encoder,
+                                     context->omx_h265_encoder,
                                      context->rtp_filter,
                                      context->rtp_queue,
-                                     context->rtp_payload,
+                                     context->rtp_h265_payload,
                                      NULL)) {
             M_ERROR("Couldn't finish pipeline linking part 4\n");
             return NULL;

bendraper

@Alex-Kushleyev Incredible, it works! Thanks Alex

Alex Kushleyev

@bendraper , great! Thank you for doing most of the work

I will add an option to specify output encoder type (for the unencoded input case) so h264/5 can be selected via command line argument.

Alex

bendraper

@Alex-Kushleyev Awesome. On a different note, do you know why ffmpeg has no issue with the aspect ratio of the stream but in QGroundControl, the stream seems to be vertically squished? I would guess ffmpeg correctly makes some assumptions about the aspect ratio that voxl-streamer doesn't outwardly declare and maybe QGC doesn't make these assumptions?

Alex Kushleyev

@bendraper , i believe that when the encoder starts producing frames, the first frame is a header frame that has the stream description. This condition is also triggered if client disconnects (encoding stops) and then reconnects, causing the stream to start again. I am not sure what the whole pipeline looks like for getting video to QGC, but if the header is not sent over or is not parsed, this can lead to incorrect stream dimension? However i would assume that is the header is lost, the video would be completely garbled, not just aspect ratio, which probably means that QGC is in fact decoding correctly but simply rendering the image using a fixed aspect ratio…

Alex Kushleyev

I just pushed two updates to a branch of voxl-streamer:

https://gitlab.com/voxl-public/voxl-sdk/services/voxl-streamer/-/commits/enable-h265-encode
add option to use h265 encoder for non-encoded input frames (use either command line option -e h265 or --encoder h265 or in voxl-streamer.conf file "encoder": "h264"
switch the h264 profile from baseline to main which allows using CABAC coding, which has better compression properties than default CAVLC at some expense of HW encoder usage (negligible) and potential additional cpu usage on decoding side. However, this will definitely be better for low bandwidth streaming. h265 uses CABAC coding as the only option.

It seems like it's working well, but not much testing has been done so far..

Alex

bendraper

@Alex-Kushleyev This is great, much appreciated! Also regarding that squished stream in QGC, seems to only be a linux thing so probably not much to do about it right now.

Eric Katzfey

@bendraper You can also set the aspect ratio of the video in QGroundControls application settings.

bendraper

@Eric-Katzfey Yea I think that functionality may be broken though. If I try to do the inverse aspect ratio, it just kinda shrinks the stream as opposed to stretching it back out

bendraper

@bendraper said in streamer default h265:

@Eric-Katzfey Yea I think that functionality may be broken though. If I try to do the inverse aspect ratio, it just kinda shrinks the stream as opposed to stretching it back out

It's also seemingly fixed in the daily builds of QGC... Oh well

bendraper

I'm looking at the camera server documentation and this is saying that H265 essentially requires twice the bandwidth that H264 does. Isn't that backwards? Isn't H265 more efficient at the cost of latency? @Alex-Kushleyev

Alex Kushleyev

@bendraper,

This is not a requirement but a current limitation of voxl-camera-server configuration of the encoder. Yes, h265 is generally 2x more efficient and having the minimum bitrate 2x compared to h264 does not make sense. This is a limitation due to some unknown mis-configuration of the encoder by the camera server. We will try to fix this, but it is not high priority, so not sure when this will happen.

The gstreamer (voxl-streamer) does not have this limitation, please double check it. I have tested voxl-streamer with bitrate down to 0.25 mbps in h265 mode and video looks surprisingly good. Such configuration can be good for streaming over very low bandwidth link.

By the way, voxl-camera-server supports dynamic bitrate updates by sending a command to camera server (within the mentioned limits), for example:

voxl-send-command hires_front_large_encoded set_large_venc_mbps 4.0

Alex

bendraper

@Alex-Kushleyev Sounds good thanks for clarifying. Also, it looks like you can use a hybrid CQP and CBR mode per the documentation, but when I use "cbr+cqp" in the camera server conf file, it does not seem to like it. Is this feature not available yet?

Alex Kushleyev

@bendraper , ah, yes the feature is available and I can see how the naming is misleading.

In order to achieve the CBR + CQP functionality, you do the following:

select cbr mode just like normal CBR
specify maximum bit rate, lets say 1.5 mbit, which is the lowest we can currently do with h264
now if you want to achieve lower bitrate when there is not much motion (instead of always forcing the desired bit rate), set _venc_Qmin to a larger number (than the default 15 or so).

Basically by default _venc_Qmin is very low, which means the encoder is allowed to increase the image quality (decrease quantization level) if there is not much motion, so that you meet the target constant bit rate.

By increasing this Qmin, you will not allow the encoder to keep increasing the quality under low motion condition, so it will cap the maximum quality level (Qmin) and reduce bandwidth. You can start with values of Qmin at 20, 25, 30 and see how low your bandwidth will get when not moving. When you start moving the camera, the bitrate should jump up to the CBR value, if needed. Please note that the quantization factor Qmin is not linear, i don't have the exact formula.

In general what happens in the encoder, if you want to maintain some CBR, the Quantization level is dynamically changed (within allowed range) depending on how much image is changing, so lots of motion in image will result in lowering overall quality (increasing quantization) in order to still be able to fit within the allowed CBR bitrate budget.

Please try it out and let me know how that works for you..

Alex

Alex Kushleyev

I have updated the docs to clarify the setting for CBR + CQP mode:

https://docs.modalai.com/voxl-camera-server/#constant-quantization-vs-constant-bitrate-mode

CBR + CQP

This is a hybrid mode that uses Constatnt Quantization when there is little motion and Constant Bitrate to limit the maximum bitrate
In order to use this mode, select cbr as the bitrate control, however also adjust the Qmin parameter, as described below