Title:Ensemble Learning for Efficient VVC Bitrate Ladder Prediction

Abstract:Changing the encoding parameters, in particular the video resolution, is a common practice before transcoding. To this end, streaming and broadcast platforms benefit from so-called bitrate ladders to determine the optimal resolution for given bitrates. However, the task of determining the bitrate ladder can usually be challenging as, on one hand, so-called fit-for-all static ladders would waste bandwidth, and on the other hand, fully specialized ladders are often not affordable in terms of computational complexity. In this paper, we propose an ML-based scheme for predicting the bitrate ladder based on the content of the video. The baseline of our solution predicts the bitrate ladder using two constituent methods, which require no encoding passes. To further enhance the performance of the constituent methods, we integrate a conditional ensemble method to aggregate their decisions, with a negligibly limited number of encoding passes. The experiment, carried out on the optimized software encoder implementation of the VVC standard, called VVenC, shows significant performance improvement. When compared to static bitrate ladder, the proposed method can offer about 13% bitrate reduction in terms of BD-BR with a negligible additional computational overhead. Conversely, when compared to the fully specialized bitrate ladder method, the proposed method can offer about 86% to 92% complexity reduction, at cost the of only 0.8% to 0.9% coding efficiency drop in terms of BD-BR.