A Rank Aggregation Framework for Video Interestingness Prediction

摘要

Often, different segments of a video may be more or less attractive for people depending on their experience in watching it. Due to this subjectiveness, the challenging task of automatically predicting whether a video segment is interesting or not has attracted a lot of attention. Current solutions are usually based on learning models trained with features from different modalities. In this paper, we propose a late fusion with rank aggregation methods for combining ranking models learned with features of different modalities and by different learning-to-rank algorithms. The experimental evaluation was conducted on a benchmarking dataset provided for the Predicting Media Interestingness Task at the MediaEval 2016. Two different modalities and four learning-to-rank algorithms are considered. The results are promising and show that the rank aggregation methods can be used to improve the overall performance, reaching gains of more than 10% over state-of-the-art solutions.