Deep reinforcement learning framework for characterizing video content

Abstract

Methods and systems for performing sequence level prediction of a video scene are described. Video information in a video scene is represented as a sequence of features depicted each frame. An environment state for each time step t corresponding to each frame is represented by the video information for time step t and predicted affective information from a previous time step t−1. An action A(t) as taken with an agent controlled by a machine learning algorithm for the frame at step t, wherein an output of the action A(t) represents affective label prediction for the frame at the time step t. A pool of predicted actions is transformed to a predicted affective history at a next time step t+1. The predictive affective history is included as part of the environment state for the next time step t+1. A reward R is generated on predicted actions up to the current time step t, by comparing them against corresponding annotated movie scene affective labels.