Influence of temporal pooling method on the objective video quality evaluation

Abstract

Objective video quality evaluation incorporates spatial and temporal degradation effects to calculate quality grade for each video frame alone, as well as the overall quality grade for the whole sequence. Distortion visibility depends not only on the level of degradation but also on video content and viewer's preferences. Temporal pooling is a method that collapses series of frame quality scores to one quality score for whole video sequence and it should be aware of an influence of the video quality variation to the human overall judgments of quality. In this paper six pooling methods applied to five objective video quality metrics are compared. The best prediction capability has been obtained by those pooling methods that reflect two phenomena: a recency effect and the worst quality section influence. Moreover, the statistics of individual frames scores tends to be of high importance and strongly influences pooling method capabilities.Index Terms-video quality, subjective assessment, objective metrics, temporal pooling

A Comparative Evaluation of Temporal Pooling Methods for Blind Video Quality Assessment

“…Simple temporal average pooling is a widely used scheme to augment both FR [11][12][13] and NR VQA models [5,7,9]. Other kinds of pooling that are used include harmonic mean [14], Minkowski mean [15,16], percentile pooling [17,18], and adaptively weighted sums [19]. More sophisticated pooling strategies have considered memory effects, such as primacy, recency [15,16,20], and hysteresis [6,10,21,22].…”

“…Other kinds of pooling that are used include harmonic mean [14], Minkowski mean [15,16], percentile pooling [17,18], and adaptively weighted sums [19]. More sophisticated pooling strategies have considered memory effects, such as primacy, recency [15,16,20], and hysteresis [6,10,21,22]. The general applicability of these pooling models, however, has not so far been deeply validated in the general context of NR VQA models for real-world UGC videos, though a few more directed studies have been conducted [15,16].…”

Design of Video Quality Metrics with Multi-Way Data Analysis

“…Rimac-Drlje et al [270] compared different pooling methods for different image quality metrics and suggest Minkowski summation over all frames, but results by You et al [372,374] indicate that Minkowski summation does not provide any advantage compared to simple averaging.…”

Quality of experience and access network traffic management of HTTP adaptive video streaming

“…A literature review found several di erent temporal pooling methods, which are investigated in this study. The most intuitive approaches described in [244] are mean pooling (Mean, 1 T • T t=1 OM (t)) and last frames mean pooling (Mean-LastFrames, 1 F • T t=T −F OM (t)). These methods simply average the objective metrics over all frames, or the F most recent frames, respectively.…”

“…1/p ) by tuning the parameter p. The additional parameter τ in ExpMinkowski controls the exponential weighting, and thus, the recency e ect. Moreover, [244] OM (t + i)]) also accounts for the poorest quality section. A related method, which was not described in [244], computes the mean of the p percent of overall frames with lowest quality (Percentile).…”