Content characterization for live video compression optimization

Compressing video sequences with different content complexity results in different bitrates for the same quality level or in different quality levels for the same bitrate; for instance it is well known that content with high spatial complexity and/or high motion requires high bitrates for compression with adequate quality. To address this, per-title optimization is used recently (e.g., by Netflix) to generate appropriate rate-quality representations for different Video on demand (VoD) content to be streamed via adaptive video streaming. However, this cannot be adopted for live video streaming as it requires encoding (multiple times) each video content. Spatial Information (SI) and Temporal Information (TI) have been often used as indicators of video complexity, for instance for preparing and describing content for video quality assessment tests, and for rate-distortion modeling. However, it has been questioned recently if different metrics could lead to a better estimation of ”compressibility” of video. In this paper we compare existing and proposed metrics in terms of their ability to estimate ”compressibility”. This supports quality-rate estimation and the possibility to create appropriate ”quality ladders” (different quality representations) for adaptive live video streaming. We observe in particular that metrics related to the variance of pixel values provide a good estimation of compressiblity for the considered datasets.