Convex Hull Prediction for Adaptive Video Steaming

#adaptive #streaming #convex #hull

Adaptive streaming technologies heavily rely on bitrate ladders that specify optimal bitrate-resolution pairs to encode video sequences. However, distinctive characteristics of video sequences meant that the bitrate ladders become content gnostic. Many works in the literature have attempted to predict these bitrate ladders without performing exhaustive trial encodes. Many such attempts however are based on a prior assumption of the range of optimal bitrates. Moreover, these methods rely on constant quantization parameter based encodings as opposed to target bitrate based encodings which are much more useful in practical applications.

In this presentation, I will elaborate on generating the optimal bitrates for a set of picture quality values across multiple resolutions (also known as convex hulls) based on the characteristics of the video contents for rate-control applications. Here, we first predict the compression complexity of video sequences using a machine learning based two-stage prediction framework. Thereafter, we derive the bitrate-resolution pairs for each target quality to generate the convex hull for that sequence. Extensive experiments demonstrate that the proposed solution outperforms state-of-the-art schemes and achieves a close approximation to the ground truth convex hull.