Lecturer: Dr. Damian Borth
This course provides an introduction to models and algorithms underlying modern multimedia data mining technology. The focus is on image and video content (though other media like text and audio will be touched as well).
For other currently offered courses, see the AGD website
Peter Burkert, Master Thesis, June 2016
Abstract: Work in progress…
Sebastian Kalkowski, Bachelor Thesis, Oct 2015
Abstract: The Yahoo Flickr Creative Commons 100 Million (YFCC100m) Dataset, has been introduced recently as the largest dataset for the computer vision and multimedia analysis community, ever created. This dataset bears a lot of potentials for computer vision and multimedia research. However its huge size and the lack of detailed knowledge about its characteristics cause accessibility problems. To increase the accessibility to the dataset, this Bachelor Thesis presents an in- depth analysis of the complete metadata included in the YFCC100m dataset and detailed insights into some of its most outstanding properties. It will be shown, that most of the metadata within the dataset has to be considered inaccurate and is generally biased. Especially the distribution of Flickr users in the dataset is strongly biased towards few highly active users, the majority of geo-locations is found within a few prominent countries and the vocabulary used for tagging favors a small set of very popular terms. Only equipped with such insights, the full potential of the YFCC100m dataset can be utilized. Additionally as a part of this thesis, the YFCC100m Browser has been developed and will be presented to the research community. This online tool is designed to grant easy access to the whole YFCC100m dataset, by implementing a search engine over the complete included textual annotations and provides real-time image and video retrieval as well as statistics about the retrieved subsets metadata. It enables users to browse the interiors of this huge dataset freely and explore interesting sub- sets of the YFCC100m collection iteratively. Additionally the unique function of the YFCC100m Browser to create, on-the-fly evaluate and finally download custom tai- lored subsets of the YFCC100m dataset - a novelty, compared to competing dataset community tools - will be presented. The design and interiors of the YFCC100m Browser will be motivated with an exemplary use-case, as well as compared with best practices in the research community for state of the art datasets, like the MS COCO, the ImageNet or MIT-Places.