Unsupervised Video Semantic Partitioning Using IBM Watson and Topic Modelling

In this paper, we present the problem of Video Semantic Partitioning that consists in breaking a video into semantically independent blocks. Defined within a framework of optimization, we present a preliminary heuristic approach to solve the problem, called Split-and-Merge. The algorithm itself is unsupervised, but the mechanisms to extract data from videos are supervised (for some) since they used IBM Watson Services. Finally, we demonstrate on few videos the capabilities of our prototype and discuss the limitations and future improvements. From the experiments, we draw two conclusions: (i) the optimal solution to the problem varies from human to human with a large variability from video to video, (ii) Split-and-Merge demonstrates encouraging qualitative results to find the average optimal solution defined as the average solution given by humans.