LinkedTV partner CERTH participated this year to the Semantic Indexing (SIN), Multimedia Event Detection (MED) and Multimedia Event Recounting (MER) tasks of the TRECVID International Benchmarking Activity. Significantly improved results were attained, compared to last year’s participation to the same tasks, verifying the significance of the work carried out in LinkedTV for solving hard video analysis problems.
One of the highlights of LinkedTV’s participation was the new, very fast machine learning method for big data problems that CERTH developed. Tested in large-scale multimedia event detection problems, this new method produces better results than both Kernel- and Linear-SVMs, while its training is one or two orders of magnitude faster than Linear-SVMs (and even more so compared to Kernel-SVMs). The use of this new approach in the Multimedia Event Detection problem was highlighted with an oral presentation at the TRECVID Workshop (check out our slides on this: http://www.iti.gr/~bmezaris/downloads/CERTH_TRECVID14_MED_Presentation.pdf ).
At the Ibero American Biennial of Design (BID) in Madrid, researchers Lilia Perez and Michiel Hildebrand from the Centrum Wiskunde & Informatica (CWI) in Amsterdam, have been awarded a honorary mention in the category Digital Design for the concept and interaction design of Linked TV News. Linked TV News is a second screen companion for watching the news based on semantic web technologies.
The project was selected out of six hundred submissions and reached the top 5 in its category. As a consequence Linked TV News will be featured in the international exhibition of the Biennial that will be shown from 25 November 2014 to 28 February 2015 in Madrid. Following that, the collection will tour throughout Spain and several Latin American countries to be shown in design and art museums and venues.
See the LinkedTV News application design video:
LinkedTV congratulates its partner CWI for this achievement!
LinkedTV partner CERTH (Thessaloniki, Greece) just publicly released version 1.4 of its software for the automatic temporal segmentation of videos into shots and scenes. This release uses exclusively CPU-based processing and the analysis is accelerated by invoking multi-threading / multi-processing operations of the Intel OpenMP runtime library, making the whole video analysis process about 7 times faster than real-time processing on an i7 PC.
The development and release of this software was supported by EU research projects LinkedTV and MediaMixer (http://www.mediamixer.eu/). Read more about research on hypervideo analysis in LinkedTV as well as check out several video analysis demonstrators.
The software is available for download at http://mklab.iti.gr/project/video-shot-segm
LinkedTV is being well represented at the International Conference on Image Processing held in Paris this year (27-30th Oct 2014). ICIP is the premier conference on Image Processing attracting this year over 1500 attendees.
LinkedTV presented its tutorial on Video Hyperlinking on the opening day of the conference. The 3 hour tutorial was well attended and generated many interesting discussions. It covered the topics of Video Analysis and HyperLink generation, building on the work carried out in LinkedTV’s workpackages on those topics led by partners CERTH and EURECOM as well as on results obtained through the participation to various international benchmarking activities (MediaEval 2014 report coming soon).
Two LinkedTV research papers on “LINKING TEXT AND VISUAL CONCEPTS SEMANTICALLY FOR CROSS MODAL MULTIMEDIA SEARCH” and “NO-REFERENCE BLUR ASSESSMENT IN NATURAL IMAGES USING FOURIER TRANSFORM AND SPATIAL PYRAMIDS” were also presented during the main conference.
Learn with LinkedTV about:
LinkedTV is pleased to contribute to the German DBPedia effort through the addition of hypernym discovery within Wikipedia article text. The “Linked Hypernyms Dataset” identifies the types of a resource through statistical type inference co-occurrence.
With this, a key problem for the German DBPedia dataset is being addressed: the sparsity of RDF type properties on DBPedia resources which makes SPARQL querying more difficult.
The new German DBPedia dataset now contains between 43 and 53% more rdf:type triples.
The same dataset is applied in the LinkedTV project to classify extracted entities from TV programming by type. This functionality is encapsulated in the service Targeted Hypernym Discovery (THD), developed and provided by our partner University of Economics Prague.
This was originally announced following the 2nd DBPedia Community meeting at http://de.dbpedia.org/node/30.