Question

I am trying to encode a video from raw YUV to MPEG-2 using the ffmpeg API.

My problem is that the API-generated file is approx. 1.7 times bigger than the equivalent files generated by ffmpeg itself.

I use the quantization parameter (via qmin and qmax) instead of the bitrate.

The API-version is basically:

//...
pCodecCtx->pix_fmt = PIX_FMT_YUV420P;
pCodecCtx->qmin = 3;
pCodecCtx->qmax = 3;
pCodecCtx->time_base.num = 1;
pCodecCtx->time_base.den = 30;
avcodec_open(pCodecCtx, avcodec_find_encoder(CODEC_ID_MPEG2VIDEO));
//...
while(/*...*/) {
  avcodec_encode_video(pCodecCtx, pOutbuf, outbufSize, pPicture);
  //..
}
//...

For ffmpeg itself, I use the command:

ffmpeg -s 352x288 -r 30 -i foreman_352x288.yuv -f mpeg2video -vcodec mpeg2video -r 30 -pix_fmt yuv420p -qmin 3 -qmax 3 foreman.m2v

Why does the API-generate file achieve a bitrate of 5212 kb/s and the file generated by ffmpeg for the same qp a bitrate of 3047 kb/s??

(Even more puzzling is that the smaller ffmpeg version has a slightly higher PSNR, 40.49 dB vs. 40.02 dB).

Are there any other relevant parameters that I missed? Does the ffmpeg actually respect the quantization parameter?

Answer 1

When using the ffmpeg API, the picture type (I-frame, P-frame, etc.) needs to be set manually for each frame. By default, ffmpeg will make every frame an I-frame.

The solution is to set the picture type before encoding a frame (here for a GOP size of 12):

//...
while(/*...*/) {
  if(pCodecCtx->frame_number % 12) 
    pPicture->pict_type = AV_PICTURE_TYPE_P;

  avcodec_encode_video(pCodecCtx, pOutbuf, outbufSize, pPicture);
  //...
}
//...

Note that setting pCodecCtx->gop_size before the encoding does not help.