29. May 2012 Dear all, About 818K new direct annotations have been produced in 2012, of which 471K were produced by TRECVID participants and about 347K were sponsored by Quaero. Combining all annotations from 2010 to 2012, about 5M direct annotations are available and turn into about 20M annotations after propagation using relations. Though the participation was only voluntary this year, 18 groups contributed significantly. Many thanks to all that participated. The final version of the 2012 collaborative annotation can be downloaded from http://mrim.imag.fr/tvca2012/ann.tgz (containing all annotations from 2010 to 2012). A raw version of the 2012 collaborative annotation can be downloaded from http://mrim.imag.fr/tvca2012/ann-raw.tgz (containing all annotations from 2010 to 2012). The raw version contains all and only the direct annotations. There may be more than one annotation per shot and concept and no synthesis (e.g. majority vote) is made. No propagation using relations is made either. The raw version is given so that participants can use whatever strategy they want for fusing the multiple annotations and for using the implies and excludes relations. The regular (full) version of the annotation is obtained from the raw one using the four steps: - propagate positive samples from specific concepts to generic concepts using the "implies" relations (e.g. positive for cat implies positive for animal); - propagate positive samples to negative samples using the "excludes" relations (e.g. positive for nighttime implies negative for daytime_outdoor); - propagate negative samples from generic concepts to specific concepts using the "implies" relations (e.g. negative for human implies negative for male_human); - take a majority decision for each annotated shot. There may be more than one annotation per concept x shot in the original (direct) annotation ; there may be even more annotations per concept x shot after the propagation takes place ; all raw annotations are propagated and the multiple are also propagated ; the majority decision is taken after all possible propagations were made. A transitive closure is implemented for the "implies" relations. We obtained better results with the full version than with the raw version but are interested in possibly better alternatives for the propagation and final decision. We provide also the raw annotations along with a set of "implies" and "excludes" relations to encourage research on this question. Improvements on this set of relations (built manually) are also welcome. Best regards, Georges. -- Dr. Georges Quénot - Chargé de Recherche CNRS Laboratoire d'Informatique de Grenoble Bâtiment B, 41 rue des Mathématiques, B.P. 53, 38041 Grenoble Cedex 9 Email: Georges.Quenot@imag.fr, Tel: (33-4) 76 63 58 55, Fax: (33-4) 76 63 56 86