KOEN
Read Original
#Backend

VMAF v1: Good Is Not Good Enough

VMAF v1: Good Is Not Good Enough
01

Summary

Netflix Unveils VMAF v1: Setting a New Gold Standard for Video Quality

From precise 4K/mobile assessment to blazing speed, explore everything about the upgraded open-source VMAF.

Netflix has overhauled VMAF, the cornerstone of its video encoding optimization, with the release of version v1. This update more precisely reflects human visual perception to maximize accuracy across diverse viewing environments while significantly boosting computational efficiency.

  • 01Precise quality prediction based on viewing distance using Barten’s CSF model
  • 02Integration of CAMBI and chroma artifact detection capabilities
  • 03Reduced computational complexity and faster execution by removing VIF
  • 04Preserving creative intent with NEG mode enabled by default
  • 05Enhanced motion features specifically for 4K and high-fps content

RECOMMENDATION

Highly recommended for streaming service operators and encoding engineers; it provides immediate benefits for platforms delivering 4K content or serving high volumes of mobile traffic.

The Problem

VMAF v0 had limitations such as preferring compression artifacts over scaling at low bitrates and difficulty generalizing across various viewing distances and devices. It also lacked sensitivity to banding and chroma artifacts and exhibited inaccuracies in evaluating high-motion or high-frame-rate (60fps) sequences.

The Solution

VMAF v1 incorporates the AIM (Additive Impairments) component into DLM and introduces distance-based modeling using Barten's CSF model. It integrates CAMBI and chroma features for better artifact detection and improves efficiency by removing the computationally expensive VIF feature while optimizing multi-threading.

The Result

The model achieved higher Spearman's rank correlation (SRCC) with subjective quality data across major datasets like WATERLOO IVC 4K. Accuracy in 4K and mobile environments improved significantly, and the tool now runs much faster due to algorithmic optimizations.

Trade-off

While overall accuracy improved, minor regressions were observed in specific datasets, and expanding the temporal window for motion features may still not fully capture the complete perceptual impact of 60fps video.

03

Key Concepts

Concept · 01

VMAF (Video Multimethod Assessment Fusion)

A video quality metric developed by Netflix that fuses multiple quality-aware features using a support-vector regressor (SVR) trained on subjective data.

  • Acts as the de facto industry standard for encoding evaluation and optimization
  • The v1 release addresses several known limitations of the previous v0 version
Concept · 02

CSF (Contrast Sensitivity Function)

A model defining human sensitivity to contrast across spatial frequencies, which relates to how distortions are perceived at different distances and resolutions.

  • VMAF v1 uses an adapted Barten’s CSF model to enable robust distance-based modeling
  • Allows the same model to be applied accurately to scenarios like phone viewing vs. 4K TV
Concept · 03

CAMBI (Contrast Aware Multiscale Banding Index)

A specialized metric designed to detect banding artifacts, which appear as staircase-like edges in smooth areas of an image.

  • Integrated into VMAF v1 as an elementary feature to capture visual quality loss due to banding
  • Optimized to ensure accuracy without introducing excessive computational overhead