Back to search

University of Waterloo

Generalized Rate-Distortion Functions of Videos

Abstract

dc:description.abstract

Customers are consuming enormous digital videos every day via various kinds of video services through terrestrial, cable, and satellite communication systems or over-the-top Internet connections. To offer the best possible services using the limited capacity of video distribution systems, these video services desire precise understanding of the relationship between the perceptual quality of a video and its media attributes, for which we term it the GRD function. In this thesis, we focus on accurately estimating the generalized rate-distortion (GRD) function with a minimal number of measurement queries. We first explore the GRD behavior of compressed digital videos in a two-dimensional space of bitrate and resolution. Our analysis on real-world GRD data reveals that all GRD functions share similar regularities, but meanwhile exhibit considerable variations across different combinations of content and encoder types. Based on the analysis, we define the theoretical space of the GRD function, which not only constructs the groundwork of the form a GRD model should take, but also determines the constraints these functions must satisfy. We propose two computational GRD models. In the first model, we assume that the quality scores are precise, and develop a robust axial-monotonic Clough-Tocher (RAMCT) interpolation method to approximate the GRD function from a moderate number of measurements. In the second model, we show that the GRD function space is a convex set residing in a Hilbert space, and that a GRD function can be estimated by solving a projection problem onto the convex set. By analyzing GRD functions that arise in practice, we approximate the infinite-dimensional theoretical space by a low-dimensional one, based on which an empirical GRD model of few parameters is proposed. To further reduce the number of queries, we present a novel sampling scheme based on a probabilistic model and an information measure. The proposed sampling method generates a sequence of queries by minimizing the overall informativeness of the remaining samples. To evaluate the performance of the GRD estimation methods, we collect a large-scale database consisting of more than $4,000$ real-world GRD functions, namely the Waterloo generalized rate-distortion (Waterloo GRD) database. Extensive comparison experiments are carried out on the database. Superiority of the two proposed GRD models over state-of-the-art approaches are attested both quantitatively and visually. Meanwhile, it is also validated that the proposed sampling algorithm consistently reduces the number of queries needed by various GRD estimation algorithms. Finally, we show the broad application scope of the proposed GRD models by exemplifying three applications: rate-distortion curve prediction, per-title encoding profile generation, and video encoder comparison.

Degree

thesis:*
Grantor dc:publisher
University of Waterloo
Year dc:date.issued
2019

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Liu, Wentao
Advisor dc:contributor.advisor
  • Wang, Zhou

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Handle dc:identifier.uri
http://hdl.handle.net/10012/14799

Chain of custody

source
Harvested from
University of Waterloo
Base URL
dspacemainprd01.lib.uwaterloo.ca/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Liu, Wentao. Generalized Rate-Distortion Functions of Videos. University of Waterloo, 2019. http://hdl.handle.net/10012/14799