Soundtrack recommendation for images
Aleksandar Stupar · Max Planck Digital Library · 2013
The drastic increase in production of multimedia content has emphasized the research concerning its organization and retrieval. In this thesis, we address the problem of music retrieval when a set of images is given as input query, i.e., the problem of soundtrack recommendation for images. The task at hand is to recommend appropriate music to be played during the presentation of a given set of query images. To tackle this problem, we formulate a hypothesis that the knowledge appropriate for the task is contained in publicly available contemporary movies. Our approach, Picasso, employs similarity search techniques inside the image and music domains, harvesting movies to form a link between the domains. To achieve a fair and unbiased comparison between different soundtrack recommendation approaches, we proposed an evaluation benchmark. The evaluation results are reported for Picasso and the baseline approach, using the proposed benchmark. We further address two efficiency aspects that arise from the Picasso approach. First, we investigate the problem of processing top-K queries with set-defined selections and propose an index structure that aims at minimizing the query answering latency. Second, we address the problem of similarity search in high-dimensional spaces and propose two enhancements to the Locality Sensitive Hashing (LSH) scheme. We also investigate the prospects of a distributed similarity search algorithm based on LSH using the MapReduce framework. Finally, we give an overview of the PicasSound|a smartphone application based on the Picasso approach.