[Libav-user] Encoding images to h264 and rtp output: SDP file without sprop-parameter-sets does not play

Rasmus rasmusdiederichsen at gmail.com
Mon Sep 13 16:57:04 EEST 2021

tl;dr: I try to encode acquired camera frames to h264, send via RTP
and play this back on another device. SDP file generated by ffmpeg for
a sample video has info which my own SDP file misses. My SDP file
plays in ffplay, but not VLC, while both play ffmpeg's SDP file. I am
suspecting missing sprop-parameter-sets in my SDP file.

Ultimately I want to play this back in VLC.


I am writing code that encodes images to h264 and outputs to an RTP
server (or client? anyway the part that is listening). I generate an
SDP file for this.
- ffplay plays the stream without problem
- mplayer shows a green box embedded in a larger black box, but I read
somewhere it only supports mpegts over RTP, so not sure
- VLC does not play the SDP file.

Now when instead I use some random video and have ffmpeg output an SDP
file like so

ffmpeg     -re     -i some.mp4     -an     -c:v copy -f rtp -sdp_file
video.sdp     "rtp://"

I can see that the generated SDP file – which plays in both ffplay and
VLC – includes the base64 encoded sprop-parameter-sets field, and
removing this causes the stream to not play.

> cat video.sdp
o=- 0 0 IN IP4
s=No Name
c=IN IP4
t=0 0
a=tool:libavformat 58.76.100
m=video 5004 RTP/AVP 96
a=rtpmap:96 H264/90000
a=fmtp:96 packetization-mode=1;

My own SDP file on the other hand, does not contain this information,
and VLC hangs for 10s and then stops trying with "no data received".

> cat test.sdp
o=- 0 0 IN IP4
s=No Name
c=IN IP4
t=0 0
a=tool:libavformat 58.76.100
m=video 44499 RTP/AVP 96
a=rtpmap:96 H264/90000
a=fmtp:96 packetization-mode=1

So my theory is that my custom code must somehow add this SPS
information to the SDP file. But despite hours of searching, I could
not find a structured way to set the extradata field on the AVStream's
AVCodeParams. The code I'm using is roughly this (I'm sure there's
unrelated errors in there):

// variables
std::vector<std::uint8_t> imgbuf;
AVFormatContext *ofmt_ctx = nullptr;
AVCodec *out_codec = nullptr;
AVStream *out_stream = nullptr;
AVCodecContext *out_codec_ctx = nullptr;
SwsContext *swsctx = nullptr;
cv::Mat canvas_;
unsigned int height_;
unsigned int width_;
unsigned int fps_;
AVFrame *frame_ = nullptr;

AVOutputFormat *format = av_guess_format("rtp", nullptr, nullptr);
const auto url = std::string("rtp://");
avformat_alloc_output_context2(ofmt_ctx, format, format->name, url.c_str());

out_codec = avcodec_find_encoder(AV_CODEC_ID_H264);
stream = avformat_new_stream(ofmt_ctx, out_codec);
out_codec_ctx = avcodec_alloc_context3(out_codec);

// not sure if these do anything, did not help in my case
out_codec_ctx->flags |= AV_CODEC_FLAG_GLOBAL_HEADER;
ofmt_ctx->flags |= AV_CODEC_FLAG_GLOBAL_HEADER;

// then, for each incoming image:
while (receive_image) {
  static bool first_time = true;
  if (first_time) {
    // discover necessary params such as image dimensions from the first
    // received image
    first_time = false;
    height_ = image.rows;
    width_ = image.cols;

    codec_ctx->codec_tag = 0;
    codec_ctx->bit_rate = 2e6;
    // does nothing, unfortunately
    codec_ctx->thread_count = 1;
    codec_ctx->codec_id = AV_CODEC_ID_H264;
    codec_ctx->codec_type = AVMEDIA_TYPE_VIDEO;
    codec_ctx->width = width_;
    codec_ctx->height = height_;
    codec_ctx->gop_size = 6;
    codec_ctx->pix_fmt = AV_PIX_FMT_YUV420P;
    codec_ctx->framerate = fps_;
    codec_ctx->time_base = av_inv_q(fps_);

    avcodec_parameters_from_context(stream, out_codec_ctx);

    // this stuff is empty: is that the problem?
    stream->codecpar->extradata = codec_ctx->extradata;
    stream->codecpar->extradata_size = codec_ctx->extradata_size;

    AVDictionary *codec_options = nullptr;
    av_dict_set(&codec_options, "profile", "high", 0);
    av_dict_set(&codec_options, "preset", "ultrafast", 0);
    av_dict_set(&codec_options, "tune", "zerolatency", 0);

    // open video encoder
    avcodec_open2(codec_ctx, codec, &codec_options);

    stream->time_base.num = 1;
    stream->time_base.den = fps_;
    avio_open(&(ofmt_ctx->pb), ofmt_ctx->filename, AVIO_FLAG_WRITE);

    /* Write a file for VLC */
    char buf[200000];
    AVFormatContext *ac[] = {ofmt_ctx};
    av_sdp_create(ac, 1, buf, 20000);
    printf("sdp:\n%s\n", buf);
    FILE *fsdp = fopen("test.sdp", "w");
    fprintf(fsdp, "%s", buf);

    swsctx = sws_getContext(width_, height_, AV_PIX_FMT_BGR24, width_, height_,
                            out_codec_ctx->pix_fmt, SWS_BICUBIC, nullptr,
                            nullptr, nullptr);

  if (!frame_) {
    frame_ = av_frame_alloc();

    std::uint8_t *framebuf = new uint8_t[av_image_get_buffer_size(
        codec_ctx->pix_fmt, width_, height_, 1)];
    av_image_fill_arrays(frame_->data, frame_->linesize, framebuf,
                         codec_ctx->pix_fmt, width, height, 1);
    frame_->width = width_;
    frame_->height = height_;
    frame_->format = static_cast<int>(codec_ctx->pix_fmt);
    success = avformat_write_header(ofmt_ctx, nullptr);
  if (imgbuf.empty()) {
    imgbuf.resize(height_ * width_ * 3 + 16);
    canvas_ = cv::Mat(height_, width_, CV_8UC3, imgbuf.data(), width_ * 3);
  } else {

  const int stride[] = {static_cast<int>(image.step[0])};

  sws_scale(swsctx, &canvas_.data, stride, 0, canvas_.rows, frame_->data,
  frame_->pts += av_rescale_q(1, out_codec_ctx->time_base, stream->time_base);

  AVPacket pkt = {0};
  avcodec_send_frame(out_codec_ctx, frame_);
  avcodec_receive_packet(out_codec_ctx, &pkt);
  av_interleaved_write_frame(ofmt_ctx, &pkt);

Can anyone offer some advice here?


More information about the Libav-user mailing list