FAUST: Fast Per-Scene Encoding Using Entropy-Based Scene Detection and Machine Learning

FAUST: Fast Per-Scene Encoding Using Entropy-Based Scene Detection and Machine Learning

30th IEEE Conference of the Open Innovations Association FRUCT
27-29 October 2021
https://www.fruct.org/

[PDF][Slides][Video]

Anatoliy Zabrovskiy, Prateek Agrawal, Christian Timmerer, and Radu Prodan

Abstract: HTTP adaptive video streaming is a widespread and sought-after technology on the Internet that allows clients to dynamically switch between different stream qualities presented in the bitrate ladder to optimize overall received video quality. Currently, there exist several approaches of different complexity for building such a ladder. The simplest method is to use a static bitrate ladder, and the more complex one is to compute a per-title encoding ladder. The main drawback of these approaches is that they do not provide bitrate ladders for scenes with different visual complexity within the video. Moreover, most modern methods require additional computationally-intensive test encodings of the entire video to construct the convex hull, used to calculate the bitrate ladder. This paper proposes a new fast per-scene encoding approach called FAUST based on 1) quick entropy-based scene detection and 2) prediction of optimized bitrate ladder for each scene using an artificial neural network. The results show that our model reduces the mean absolute error to 0.15, the mean square error to 0.08, and the bitrate to 13.5% while increasing the difference in video multimethod assessment fusion to 5.6 points.

This entry was posted in APOLLO, ATHENA. Bookmark the permalink.