@InProceedings{Kletz2018, title = {Evaluation of Visual Content Descriptors for Supporting Ad-Hoc Video Search Tasks at the Video Browser Showdown}, author = {Sabrina Kletz and Andreas Leibetseder and Klaus Schöffmann}, booktitle = {MultiMedia Modeling - 24th International Conference, MMM 2018 (Part 1)}, year = {2018}, address = {Berlin}, editor = {Klaus Schöffmann and Thanarat H. Chalidabhongse and Chong-Wah Ngo and Noel E. O´Connor and Supavadee Aramvith and Yo-Sung Ho and Moncef Gabbouj and Ahmed Elgammal}, month = {Januar}, pages = {203--215}, publisher = {Springer}, series = {LNCS}, volume = {10704}, abstract = {Since 2017 the Video Browser Showdown (VBS) collaborates with TRECVID and interactively evaluates Ad-Hoc Video Search (AVS) tasks, in addition to Known-Item Search (KIS) tasks. In this video search competition the participants have to find relevant target scenes to a given textual query within a specific time limit, in a large dataset consisting of 600 h of video content. Since usually the number of relevant scenes for such an AVS query is rather high, the teams at the VBS 2017 could find only a small portion of them. One way to support them at the interactive search would be to automatically retrieve other similar instances of an already found target scene. However, it is unclear which content descriptors should be used for such an automatic video content search, using a query-by-example approach. Therefore, in this paper we investigate several different visual content descriptors (CNN Features, CEDD, COMO, HOG, Feature Signatures and HOF) for the purpose of similarity search in the TRECVID IACC.3 dataset, used for the VBS. Our evaluation shows that there is no single descriptor that works best for every AVS query, however, when considering the total performance over all 30 AVS tasks of TRECVID 2016, CNN features provide the best performance.}, doi = {10.1007/978-3-319-73603-7_17}, url = {https://link.springer.com/chapter/10.1007/978-3-319-73603-7_17} }