CERTH @ MediaEval 2012 SED Tool
Social Event Detection (SED) in tagged photo collections
Description
This is a Java library implementing the framework presented by CERTH in the Social Event Detection task at MediaEval 2012. It can be used as a competing method for detecting events in large tagged photo collections.
Instructions
1. Download the library and auxiliary data tar.gz archives. The library jar file contains all necessary dependencies (jars).
2. Preprocessing step: The preprocessing step aims to create the lists of candidate items for each item in the dataset based on time, location and visual and textual features. These candidates have already been computed and are placed under the candidates_per_item folder (zipped). Also in this step two global visual features are computed. Specifically for each image of the dataset a gist and vlad+surf descriptors are computed and stored into appropriate data structures (visual_features folder).
3. Learn same class model and create candidate events: This step may require a lot of time to be accomplished.
Command: java -jar sed.jar mediaeval.MediaEval </path/to/config.xml>
In this step you have to specify the following set of parameters:
- training_set_folder: the folder contains the essentials for the training procedure
- metadata_file: the xml file contains the metadata for each image of the dataset
- codebook_file: a file contains the textual codebook, computed by the sed's dataset
- visual_features_folder: the folder that contains the visual features (gist and vlad+surf)
- candidates_per_item_folder: the folder that contains the nearest neighbors for each item in the dataset according to the various modalities
- candidate_events_file: the output file which contains the candidate events (one event per line)
4. Post processing: Merging and splitting of events based on time and location. Also, filtering of irrelevant events.
Command: java -jar sed.jar event.Events </path/to/config.xml>
In this step you have to specify the following set of parameters:
- jaccard_similarity_threshold: similarity threshold beetwen an event and a concept
- number_of_dominant_terms: number of dominant terms in the concept
- output_folder: the output folder of the procedure
- concept_terms_file: the file that contains terms relevant to the specific concept
- general_terms_file: the file that contains a set of general terms challenge: 1, 2 or 3
- splitting_time: split events that exceeds this time interval
- include_single_items: true or false if we want to inlude single items as events
- merge_events: true or false if we want to merge clusters of items into larger events
- merging_time: merge events according to this time distance
- bboxies: an array of bounding boxes used to filter out events irrelevant to these locations
- gt_folder: the folder that contains the ground truth files
Publication
If you make use of this software for your research work, please cite the following paper:
E. Schinas, G. Petkos, S. Papadopoulos, Y. Kompatsiaris, “CERTH @ MediaEval 2012 Social Event Detection Task” In Proceedings of MediaEval 2012 Workshop, Pisa, Italy, 4-5 October 2012
License
THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Acknowledgements
This work was supported by the EU FP7 project SocialSensor (FP7-287975).
Contact
You may contact Manos Schinas (manosetro AT iti DOT gr) and Symeon Papadopoulos (papadop AT iti DOT gr) for any question or remark you may have with respect to this tool.