LIVE-FB LSVQ (LIVE-FB Large-Scale Social Video Quality (LSVQ) Database)

无参考 (NR) 感知视频质量评估 (VQA) 是社交和流媒体应用的一个复杂、未解决且重要的问题。需要高效和准确的视频质量预测器来监控和指导数十亿共享的、通常不完美的用户生成内容 (UGC) 的处理。不幸的是，当前的 NR 模型对真实世界的“野外”UGC 视频数据的预测能力有限。为了推进这个问题的进展，我们创建了（迄今为止）最大的主观视频质量数据集，包含 39,000 个真实世界失真视频和 117,000 个时空本地化视频补丁（“v-patches”），以及 5.5M人类感知质量注释。使用它，我们创建了两个独特的 NR-VQA 模型：（a）一个基于局部到全局区域的 NR VQA 架构（称为 PVQ），它学习预测全局视频质量并在 3 UGC 数据集，以及 (b) 首创的时空视频质量映射引擎（称为 PVQ Mapper），有助于定位和可视化空间和时间的感知失真。我们将在审查过程后立即提供新的数据库和预测模型。

No-Reference (NR) Perceptual Video Quality Assessment (VQA) is a complex, unsolved, and critical problem in social and streaming applications. Efficient and accurate video quality predictors are required to monitor and guide the processing of billions of shared, often imperfect user-generated content (UGC). Unfortunately, current NR models have limited predictive performance on real-world "in-the-wild" UGC video datasets. To advance progress on this problem, we present the largest subjective video quality dataset to date, which includes 39,000 real-world distorted videos, 117,000 spatially and temporally localized video patches ("v-patches"), and 5.5 million human perceptual quality annotations. Using this dataset, we develop two distinct NR-VQA models: (a) a local-to-global region-based NR VQA architecture named PVQ, which learns to predict global video quality and performs well across 3 UGC datasets; and (b) the first-of-its-kind spatiotemporal video quality mapping engine dubbed PVQ Mapper, which enables the localization and visualization of spatial and temporal perceptual distortions. We will make the newly created database and prediction models publicly available immediately after the review process.

搜集汇总

背景与挑战

背景概述

LIVE-FB LSVQ是一个大规模的视频质量评估数据集，包含大量真实世界失真视频和人类感知质量注释，旨在推动无参考视频质量评估模型的发展。