Near Duplicate Detection

This is a prototype system that supports reverse video search on a large collection of web videos hosted by ITI-CERTH. In our four-year research on video retrieval, we have developed a two-stage algorithm for the fast and accurate video search. Given a query video to the system, an efficient indexing scheme[1] takes place in the first stage, in order to narrow down the videos in the “index” to a set of candidates whose similarity to the query has to be calculated. In the second stage, an accurate video similarity learning[2] network assesses the similarity between the candidate videos and the query, considering the spatial and temporal similarity structure of the compared videos.

The use of search operation is straightforward: you need to enter the URL of an input video in the “search” box, and the system retrieves from its “index” a set of highly similar videos in terms of visual appearance. Similar videos often correspond to near-duplicate or partial duplicate versions of the input video, while there are also cases where related videos from the same event are retrieved, and even irrelevant videos may occasionally pop up. We would appreciate your feedback in case you run into irrelevant or surprising results. Feel free to get in touch with the email below.

The video “index” for this prototype system is the FIVR-200K dataset[3], a collection of more than 200 thousand videos from YouTube that were posted between the beginning of 2013 and end of 2017 and are related to queries related to natural disasters, accidents and war events crawled from Wikipedia's Current Event page. A more detailed analysis of the data collection and annotation process is available in the above link. In addition to this fixed video collection, users of this system can extend the index by issuing new queries to YouTube in the “Add” page. Note that adding videos to the index is a relatively time-consuming process, because several queries need to be issued to YouTube, and then the videos need to be analysed.

This prototype system has been developed with the support of the Horizon 2020 InVID and WeVerify projects. The back-end service and video retrieval algorithm have been developed by Giorgos Kordopatis-Zilos, while the front-end was developed by Lazaros Apostolidis. Contributions in terms of concept and operational design were made by Symeon Papadopoulos. In case you have any questions regarding the prototype system and the underlying research, please get in touch with us.

[1] G. Kordopatis-Zilos, S. Papadopoulos, I. Patras, and I. Kompatsiaris. Near-duplicate video retrieval by aggregating intermediate cnn layers. In Proceedings of the international conference on Multimedia Modeling, 2017.
[2] G. Kordopatis-Zilos, S. Papadopoulos, I. Patras, and I. Kompatsiaris. ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning. In Proceedings of the IEEE International Conference on Computer Vision, 2019.
[3] G. Kordopatis-Zilos, S. Papadopoulos, I. Patras, and I. Kompatsiaris. FIVR: Fine-grained Incident Video Retrieval. IEEE Transactions on Multimedia, 2019.