Impression Estimation of Video and Application to Video Creation

Abstract

Adding BGM (background music) to a video is an important process in video creation because BGM determines the impression of the video. We model impression estimation of a video as mappping from computer-mesurable audio and visual features to impression degrees. As an application of impression estimation of a video, we propose OtoPittan, a system for recommending BGM for helping users to make impressive videos. OtoPittan regards the problem of selecting BGM from a music collection as a partial inverse problem of the impression estimation. That is, to an inputted video and desired impression, BGM which produces a good match to the desired impression when adding it to the inputted video is recommended. As implementation ways of impression estimation of a video, we use a static user model and a dynamic user model. The first model statically constructs a mapping function learnt from training data. The second model dynamically optimizes a mapping function through user interaction. Experimental results have shown that the static user model has high estimation accuracy and the dynamic user model can efficiently performs optimization without much user interaction.