Rate-Distortion Optimized Temporal Segmentation Using Reinforcement Learning for Video Coding

In this paper, we present a reinforcement learning (RL)-based coding method to recursively divide video frames into several groups displaying similar temporal characteristics and improve rate-distortion (R-D) performance. Although the previous works have attempted to challenge the problem with analytical models, it was difficult to address complicated de-pendencies of video frames. In the proposed method, we cast the recursive problem as a sequence of a state-action for an agent to conduct an RL, by partitioning the current group to the half. The optimal solution is obtained by maximizing a reward function of the RL policy. Experimental results demonstrate that the proposed method can adapt to a video sequence whereas a fixed coding scheme cannot efficiently achieve optimal coding performance in dynamic video sequences.