After the dinner cruise along the river Spree, the second day of Adaptive Multimedia Retrieval 2008 again starts with an interesting invited talk on the European answer to Google search engine technology - THESEUS.
Karsten Müller from Fraunhofer Heinrich-Hertz-Institute is presenting on "THESEUS Project - Applications and Core Technologies for the Semantic Web". First, Karsten makes clear, that THESEUS doesn't want to be Google ;-) THESEUS is a research program for a new internetbased knowledge infrastructure....which from my point of view means nothing else but "the semantic web"....
One part of the THESEUS project is ALEXANDRIA, the virtual library, being lead by Yahoo! with the objective of semantic processing of different forms of content to enable faster access to relevant content, which again means an increase in information quality. Concepts such as an automated tagging framework (including language error correction, synonym & tag merging, and topic focussing, identification of semantic relations), innovative navigation (by presenting thematically related contents) and interaction concepts are involved.
Another part is ORDO, which deals with "Organizing your digital life" with the goal to unify various data formats, multilingual information, structured and unstructured data on the web to enable homogeneous information sources.Problems such as separating important from unimportant, ordering information instead of searching, priorization, identification and visualization of interrelations are addressed.
TEXO is another part with the objective of "Realizing the internet of services" (being lead by SAP Research), offering personalized customized services, community involvement to improve services, as well as a smooth & seamless (userfriendly) adaption and integration of services.
PROCESSUS deals with the "Optimization of business processes" aiming for the objective of anytime providing the user with theright information at any stage of the business process.
MEDICO is another subproject dealing with "Towards Scalable Semantic Image Search in Medicine" and being lead by Siemens.
CONTENTUS, as being the last Use case "Content access and generation from cultural institutions is lead by the Deutsche Nationalbibliothek. Being part of CONTENTUS are tasks such as Digitizing books as well as audiovisual material (including the German Music Archive in Berlin) protecting the cultural heritage. The goal is the semantically interlinked collection of content to achieve a next generation multimedia library.
.....impressive and ambitious project!
The upcoming section this morning is on "Image Tagging" and Marius Renn (at least I hope so) from TU Kaiserslautern is givig a presentation on "Automatic Image Tagging using Community-Driven Online Image Database". Automatic image tagging requires a lot of training data....and flickr is delivering tons of tags per day...but are these flickr data really good candidates for learning? So, in the end, unfiltered community image sets directly do not provide satisfying results. Alas, these databases at least allow large scale image aggregation...
The next talk in this session is given by Christian Hentschel from Fraunhofer HHI Berlin about "Automatic Image Annotation Refinement using Object Co-Occurences". Again, flickr is the target image set with its huge collection of more than 2 billion images, growing by 3 million photos every day. Objects always appear and are perceived in a semantic context.
The following session is on "Symbolic Music Retrieval" and starts with Rainer Typke from Austrian Research Institute for Artificial Intelligence (ÖFAI), but I had to skip this talk. Anyway, the samples of the reduced MIDI files were quite interesting (although I'm not a fan of the Scorpions!). OK, I had to ask afterwards about the usefulness and application of his approach. In music retrieval it can be used to reduce the index size down to 30% of the original index. Also QBE-processing will become much easier while on the other hand you might connect this MIDI-collection to real music files.
The last talk of the morning session is given by Giancarlo Vercellesi from University of Milan on "Automatic synchronization between audio and partial music score presentation". He presents the ParSi architecture, which perfoms an alignment of PCM signal and partial MIDI scores.
The afternoon session is simply entitled with "Systems". Fernando Lopéz from Madrid is giving a presentation on "Towards a fully MPEG-21 compliant adaption engine: complementary description tools and architectural models". Within the MPEG-21 framework several aspects of metadata-driven adaption is not clearly covered. He introduces CAIN, a tool for adapting Digital Items e.g. to different output devices.
The session continues with a presentation on "Mobile museum guide based on fast SIFT recognition" with the objective to identify paintings in galleries simply with the help of mobile pattern recognition without any extra installation on site. The SIFT (Scale Invariant Feature Transform) method is a rather cool algorithm for detecting local features within images that are used to map photographs taken with your PDA or mobile phone in the image gallery with reference pictures from a given database. And actually the live demo did work :)
I guess, we will also use the SIFT-algorithm in yovisto for synchronization of ppt/pdf-slides with the lecture video.
For the last session - "Structuring of Image Collections" - only one speaker showed up. Marc Gelgon is presenting on "Geo-temporal structuring of a personal image database with two-level variational Bayes mixture estimation".
[to be continued...]