Skip to content

Research-0652: CHUG Visual-Signal Primitives

Question

Which unblocked signal gap from the fresh signal-mix audit can be closed without waiting for a trained HDR model or a new public dataset?

Findings

  • CHUG feature extraction already decodes matched reference and distorted clips to aligned yuv420p10le files before running libvmaf.
  • The CHUG MOS head needs row-local cues for blur/noise/grain experiments. Canonical six libvmaf features are not explicit about those axes.
  • A small sampled luma pass can produce deterministic proxies: luma standard deviation, Laplacian variance, neighboring-pixel high-frequency energy, and robust Laplacian MAD.
  • These primitives are cheap enough to compute during materialisation and can be cached next to the existing libvmaf feature cache.

Decision Input

Add feature_ref_*, feature_dis_*, and feature_delta_* columns for:

  • luma_std
  • sharpness_laplacian_var
  • highfreq_abs_mean
  • noise_lap_mad

The names are deliberately descriptive rather than claiming a full perceptual model. In particular, high Laplacian variance may represent detail or noise; downstream audits must inspect it jointly with MOS and other features.

Validation Plan

  • Unit-test the YUV10 reader and signal calculator against a synthetic checker pattern.
  • Unit-test CHUG materialisation so the emitted row includes reference, distorted, and delta visual fields.
  • Keep the real ffmpeg/vmaf degraded-pair smoke as a separate optional command, with VMAF_BIN_FOR_TESTS pinned to the known-good build when the stale host binary stalls.