SENSEI: Aligning Video Streaming Quality with Dynamic User Sensitivity

https://people.cs.uchicago.edu/~junchenj/docs/sensei.pdf

Talk

• Network bandwidth is insufficient for desirable QoE

• Goal: better QoE for more users given limited bandwidth

• Conventional wisdom

  • Treat video chunks equally when the player choose bitrate for chunks

  • Key insight: users have different quality sensitivity to the chunks

• Different quality tolerance to rebuffering

• Quality sensitivity varies with video content! (different degree of attention to different parts of videos)

• Roadmap

  • Demonstrate high variability of quality sensitivity in real videos

    • 

    • QoE drop could vary > 110% for 50% videos!

    • Opportunity: large variability enables us to trade off insensitive chunks for sensitive ones

  • Quantify this quality sensitivity reliably

  • Leverage this quality sensitivity to improve adaptive video streaming

• Incorporating quality sensitivity into a QoE model

  • Traditional QoE model: sum of QoE estimate of individual chunks

  • In SENSEI: reweight the chunks by their quality sensitivity in a QoE model

  • How to capture content-dependent quality sensitivity?

    • Strawman: directly use video saliency models

      • Pixel-motion-based models, e.g., AMVM

      • Interestingness score models, e.g., Video2Gif, DSN

      • However, the purposes of the saliency models do not align with quality sensitivity

• Idea: directly ask for quality sensitivity by crowdsourcing

  • Pros

    • Directly link video quality to QoE

    • Worth the cost for popular on-demand videos

  • Cons

    • High cost to evaluate every chunk and every type of low-quality event

      • Idea: coarse quality sensitivity

        • Group chunks that might have similar quality sensitivity

        • Zoom in the representative chunks in each group

      • Two step scheduling

        • Step 1: identify chunks that share weights

        • Step 2: zoom in the representative chunks to get the weight

    • Response reliability affecting the QoE model accuracy

      • Challenge: crowdsourcing workers might provide random responses

      • Quality control scheme:

        • Engagement test, control questions, randomized video order, use Master Turkers

        • More reliable responses make higher accuracy of the QoE model

    • Not support live-video streaming

• Protect quality sensitive video chunks

  • New action: lower the quality of insensitivity chunks to get high quality for sensitive chunks

  • 

• Evaluation

  • Dataset: 16 videos

  • Baseline ABR algorithms: Fugu, Pensive, BBA (buffer-based)

  • Sensei achieves higher QoE

  • Sensei can save bandwidth

• Summary

  • Observation: for viewers, quality sensitivity varies as video content changes

  • Key idea: embrace variability of quality sensitivity by sensitivity weights obtained via per-video crowdsourcing

  • SENSEI improves video QoE by 15.1% or save bandwidth by 26.8% on average with a cost of $31.4 per min video

Some take-aways and questions:

• The key insight is the content-dependent dynamic quality sensitivity

• Approach: separate crowdsourcing experiment for each video to derive the quality sensitivity of users at different parts of the video

  • Dynamically align higher (lower) quality with higher (lower) sensitivity period

• Compare to prior works

  • Recent adaptive bitrate (ABR) algorithms: near-optimal balance between bitrate and rebuffering events

  • Recent video codecs: improve encoding efficiency but require more compute power

  • New trends: better tradeoffs between bandwidth usage and user-perceived QoE