v1.2.0

Added

• qc sample: extracts a sample of one or several videos into one
  file, each video giving the same length of scenes chosen on its spatial
  and temporal information: the most complex (--scenes top), a mix of
  complex and representative ones (mixed, the default, --top-share),
  representative ones (average) or the easiest (easy); --duration
  sets its length. The sample plays the most complex scene first (the
  easiest first for easy). The scenes are whole GOPs copied from the sources
  without re-encoding, so the frames are the sources' own; the sample is
  then read back (frame count, timestamps, a decode of every frame).
  Library: package sample, encode.FFmpeg.Copy and Decodes. The wizard
  offers it for one video or several.

• Banding inherited from the source: when an encode has banded frames,
  CAMBI is measured on the reference at those frames (cambi_source,
  banding.sourceFrames), in a second pass over them alone. Comparisons
  say how many banded frames the reference has too, and a ladder rung whose
  banding is mostly the source's is reported as such (banded-source)
  instead of being advised a 10-bit encode.

• Harmonic mean and drops in sampled mode: the harmonic mean is now
  estimated from the sampled clips, and every measurement reports the share
  of frames scoring more than 10 VMAF under the mean (drops), a finding
  from 5% of the frames.

• A note on ladders of HDR sources built on the PQ or HLG signal: their
  VMAF targets rank the encodes but are not calibrated (hdr-ladder-target).

Changed

• CAMBI is told the encode's resolution and bit depth
  (cambi.enc_width, enc_height, enc_bitdepth), as Netflix recommends
  for VMAF v1 (vmaf.ScorerConfig.Encoded). Renditions below the evaluation
  resolution read less banding (CAMBI about halved at 360p and 540p) and
  0.2 to 0.9 VMAF more; renditions at the evaluation resolution are
  unchanged. Scores of lower rungs are therefore not comparable with those
  of 1.1.0.
• The 6 VMAF step between rungs is no longer described as "one
  just-noticeable difference": measured ones average about 7 and vary with
  the content. The default is unchanged.

Documentation

• What the confidence interval is (sampling error) and is not (the
  uncertainty of VMAF itself); the sampler's coverage on SVT-AV1 encodes;
  the error of reading the codec comparison between rungs.

Docker

docker run --rm -v "$PWD:/data" ghcr.io/eko/qc:1.2.0 version --check

Binaries

Self-contained (libvmaf and the VMAF models are built in); ffmpeg is still
needed. Linux binaries are static; the macOS one is not signed
(install).