Low Complexity "Smart" Per-Scene Video Encoding

Recently, Netflix proposed a perceptual per-scene video encoding framework for over-the-top (OTT) adaptive streaming applications. Given a full-matrix of fixed quantization (fixed-QP) scene encodes (i.e., all scenes encoded at all QP values for all resolutions), it selects scene encodes which reduce the average bitrate for a target quality value. This paper proposes a low complexity alternative for achieving per-scene encoding. At first, using a small set of scene encodes, a bitrate-resolution-quality model of the complete video is constructed. Using this model, per-scene video encoding is posed as a global optimization problem. The proposed method adapts the instantaneous bitrate and encoding resolution to the complexity of the scene, enabling lower overall bitrate compared to the adaptive-QP approach [2]. Additionally, thanks to our bitrate-resolution-quality modeling, the proposed method significantly reduces the computational complexity without the need to compute a full-matrix of scene encodes and their respective quality computations.