# Video datasets

For videos *with* audio tracks, we recommend encoding videos using `ffmpeg` with the following
settings:


```bash
ffmpeg -i INPUT \
  -preset veryslow \
  -keyint_min 2 -g 24 -sc_threshold 0 \
  -c:v libx264 -pix_fmt yuv420p -crf 17 \
  -c:a aac -b:a 256k -ac 2 -ar 48000 \
  -movflags +frag_keyframe+empty_moov+default_base_moof \
  OUTPUT.mp4
```

This ensures wide browser compatibility. A CRF value below 18 is generally considered [visually
lossless](https://trac.ffmpeg.org/wiki/Encode/H.264). For audio, compression at 256 kbps
or above is considered unnoticeable [[1, 2](#references)].

For videos *without* audio, use the following settings:


```bash
ffmpeg -i INPUT \
  -preset veryslow \
  -keyint_min 2 -g 24 -sc_threshold 0 \
  -c:v libx264 -pix_fmt yuv420p -crf 17 \
  -an \
  -movflags +frag_keyframe+empty_moov+default_base_moof \
  OUTPUT.mp4
```

The `-movflags` option creates fragmented mp4 files which are required in some of our
experiments. If videos are not fragmented, our platform will automatically fragment videos
using `ffmpeg`, but will not transcode videos.

#### References

[1] Pras et al. (2009). Subjective Evaluation of MP3 Compression for Different Musical Genres.
[2] Hines et al. (2014). Perceived Audio Quality for Streaming Stereo Music.