Convex Hull Prediction for Adaptive Video Streaming by Recurrent Learning.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Adaptive video streaming requires efficient bitrate ladders for optimal quality under bandwidth limitations.
• Traditional methods involve computationally intensive pre-encoding to determine optimal rate-quality curves (convex hulls).
• This pre-encoding process leads to significant time and computational overhead.

Purpose of the Study:

• To propose a novel deep learning-based method for predicting convex hulls of video content.
• To reduce the computational and time overhead associated with traditional bitrate ladder selection.
• To enhance the efficiency of adaptive video streaming.

Main Methods:

• A recurrent convolutional network (RCN) was developed to analyze video spatiotemporal complexity.
• The RCN implicitly predicts convex hulls for video shots.
• A two-step transfer learning scheme was employed for model training, ensuring content diversity and capturing source video statistics.

Main Results:

• The proposed RCN-Hull model achieved better approximations of optimal convex hulls compared to existing methods.
• Significant time savings were observed, with an average reduction of 53.8% in pre-encoding time.
• The predicted convex hulls showed minimal deviation from ground truth, with an average Bjøntegaard delta bitrate (BD-rate) of 0.26% and a mean absolute deviation of 0.57%.

Conclusions:

• The deep learning approach effectively predicts convex hulls for adaptive video streaming.
• The method offers substantial computational and time savings without compromising video quality.
• This advancement contributes to more efficient and effective adaptive video streaming solutions.