WWDC Quick Look 💓 By SwiftGGTeam
中文 EN 日本語
What's new in AVQT

What's new in AVQT

VideoAVQTQualityEncoding
Watch original video

Highlight

AVQT (Advanced Video Quality Tool) brings three key updates at WWDC 2022: interactive HTML visual reporting, time window selection to support precise evaluation of specific clips, and Linux platform support (beta version). At the same time, the raw YUV format supports expansion from a few to 20 types.

Core Content

Why AVQT is needed

After the video is encoded, how do you know whether the quality is good or not? Viewing with the naked eye is too subjective, and objective indicators such as PSNR deviate from people’s true feelings. AVQT is a full-reference video quality assessment tool that compares the original video and the compressed/processed video and outputs a quality score that is close to human subjective ratings.

（00:26）

AVQT has three unique advantages:

  • High Correlation: The scores are highly consistent with people’s subjective ratings, suitable for various content types such as animation, natural scenes, sports, etc.
  • Fast: Evaluate a 2-hour 4K HEVC movie in just 20 minutes on M1 Ultra (6x real-time speed) based on AVFoundation and Metal
  • Configurable: Support different viewing scenarios, adjust evaluation parameters through display resolution, size and viewing distance

Interactive HTML report

Previously, AVQT only output command line numbers, making it difficult to visually judge the distribution of quality on the timeline. The new version adds--visualizeParameters to generate an interactive HTML report.

（02:33）

The report contains:

  • Frame-by-frame and segmented AVQT score graphs, supporting zooming and panning
  • Display specific time, score and frame index on mouseover
  • Pie chart showing the distribution of scores in five quality levels (Bad, Poor, Fair, Good, Excellent)
  • PSNR score comparison included at the bottom

The generated reports are pure HTML files that can be opened directly in Safari without the need to install AVQT. Facilitate sharing and discussion among team members.

Time window evaluation

An average score for the entire video may mask local issues. Take a 10-minute video, for example, with 30 seconds showing severe compression artifacts, but the overall average score is still normal.

（05:03）

The new version supports four new parameters to precisely specify the evaluation range:

  • --start-frame / --end-frame: Specify by frame number ---start-time / --end-time: Specified by timestamp

This allows for direct evaluation of specific scenes, transitions, or high-complexity clips faster than processing the entire video and filtering it manually.

Raw YUV format extension

AVQT originally supports all AVFoundation video formats, and also supports raw YUV. In 2022, raw YUV support will be expanded from a few to 20 types, covering multiple chroma sampling (4:2:0, 4:2:2, 4:4:4) and bit depth (8/10/12/16-bit).

（06:37）

The parameter system has also been adjusted: the oldreference-fourccandtest-fourccDeprecated, changed tochroma-subsamplingandbit-depthtwo independent parameters. The new parameters are clearer and easier to extend.

Linux version (Beta)

Many video encoding pipelines are deployed on Linux servers or clouds. Previously, the video needed to be transferred back to Mac to run AVQT, which increased latency and complexity.

（07:31）

AVQT for Linux (beta) allows developers to evaluate video quality directly on a Linux server. Features:

  • No external dependencies required, plug and play
  • The command line parameters are exactly the same as the macOS version
  • Supports all 20 raw YUV formats
  • The output file format is the same as the macOS version
  • Viewing condition parameters are currently unavailable and will be added in subsequent versions.

Detailed Content

Basic usage

Evaluate the quality of a compressed video:

avqt -reference reference.mov -test compressed.mov

Output example:

AVQT Score: 0.85
Processing time: 18m 32s

Key points:

  • -referenceSpecify original lossless video --testSpecify the compressed/processed video to be evaluated
  • The resolution and duration of the two videos must match
  • Score range 0-1, the closer to 1 the better the quality

Generate interactive reports

avqt -reference reference.mov -test compressed.mov --visualize

After the operation is completed, the terminal will output the report file path:

Report generated: /path/to/report.html

Open in Safari to view the interactive chart.

Key points:

  • --visualizeParameters trigger HTML report generation
  • Reports include interactive frame-by-frame quality score curves
  • Supports zooming and panning to check any time period
  • Pie chart showing distribution of quality levels
  • Can be shared directly with colleagues without installing AVQT

Evaluate specific scenarios

Use QuickTime Player to find the start and end frame numbers of the target scene, then:

avqt -reference reference.mov \
     -test compressed.mov \
     --start-frame 270 \
     --end-frame 486

Key points:

  • --start-frameand--end-framePrecisely crop the evaluation range by frame number
  • Only process specified fragments, faster than full film evaluation
  • Suitable for comparing the performance of different encoding parameters in specific scenarios
  • Also available--start-timeand--end-timeSpecify by timestamp

Processing Raw YUV video

avqt -reference reference.yuv \
     -test compressed.yuv \
     -width 1920 \
     -height 1080 \
     --chroma-subsampling 4:2:0 \
     --bit-depth 10

Key points:

  • Raw YUV requires manual specification of width and height ---chroma-subsamplingSpecify chroma sampling format ---bit-depthSpecify bit depth
  • old parametersreference-fourccandtest-fourccDeprecated and should be migrated to new parameters

Use on Linux server

# Download AVQT for Linux and extract it
./avqt -reference input.mov -test output.mov --visualize

Key points:

  • The Linux version is beta, and the command line interface is exactly the same as macOS
  • No need to install additional dependencies
  • Suitable for integration into CI/CD coding pipelines
  • Viewing condition parameters (display resolution, size, distance) are temporarily unavailable

Core Takeaways

Build an automated video quality monitoring pipeline

  • What to do: Automatically run AVQT after the video encoding is completed, and automatically adjust the encoding parameters and try again when the quality is not up to standard.
  • Why it’s worth it: The Linux version allows AVQT to be deployed on encoding servers, combined with--visualizeGenerate reports for your team to review
  • How to start: Add an AVQT evaluation step after the encoding script, set a score threshold (such as 0.8), and increase the code rate to re-encode when it is below the threshold.

Build video quality regression testing

  • What to do: Maintain a set of representative test clips and run AVQT to compare score changes after each encoder upgrade
  • Why it’s worth it: The time window feature allows you to focus on scene types where the encoder is weakest (e.g. dark areas, fast motion)
  • How to start: Pick 5-10 typical scenarios and use--start-frame/--end-frameAccurate assessment, log scores to database to track trends

Video compression parameter tuning tool

  • What: A visualization tool to batch test different encoding parameter combinations and use AVQT to report visual comparisons of quality differences
  • Why it’s worth it: Interactive HTML reporting allows non-technical people to understand quality differences and accelerate decision-making
  • How to start: Use a script to batch generate versions with different parameters and run AVQT--visualize, aggregate the reports into a comparison page

Real-time quality monitoring for live streaming platforms

  • What to do: Regularly intercept clips on the live streaming side, use AVQT to evaluate the encoding quality, and promptly detect the degradation of image quality
  • Why it’s worth it: AVQT 6x real-time on M1 Ultra, fast enough for near real-time evaluation
  • How to start: Intercept a comparison clip between the reference source and the encoding output every few minutes, run AVQT in the background, and alert when the score is abnormal.

Comments

GitHub Issues · utterances