Content-based Cluster Analysis and Cluster Summary Generation (Master Thesis, Finished)


Damian Knuchel


Exploring large multimedia collections containing several thousand, if not hundreds of thousands of items, as is often the case even with personal image libraries, is nearly impossible for a single person. One way to make such collections accessible for browsing without prior knowledge of what it contains is to perform cluster analysis, which groups items based on proximity. Visual-text co-embeddings are currently very popular features for multimedia retrieval due to their versatility and ability to encode content-based semantic information in a vector space. By performing cluster analysis on such co-embedding spaces it is possible to reveal the underlying structure of multimedia collections. Provided such a clustering, it may be possible choose representative items from each cluster or even generate cluster descriptions through automatic captioning, allowing for the summarization or extraction of information from these clusters, helping to provide an overview of the collection.

The goal of this project is to implement and evaluate different cluster analysis methods as well as methods to summarize these clusters.

Start / End Dates

2023/05/01 - 2023/10/31


Research Topics