CERTH's participation at the photo annotation task of ImageCLEF 2012
Eleni Mantziou, Georgios Petkos, Symeon Papadopoulos, Christos Sagonas, Ioannis Yiannis Kompatsiaris · 2012
Abstract. This paper describes the approaches and experimental settings of the five runs submitted by CERTH at the photo annotation task of ImageCLEF 2012. Two different approaches were used, the first using the Laplacian Eigenmaps of an image similarity graph for learning, and the second using a “same class ” learning model. Four runs were submitted using the first, and one using the second approach. A multitude of textual and visual features were employed, making use of different aggregation (BoW, VLAD) and post-processing schemes (WordNet, pLSA). The best performance scores in the test set was achieved by Run 3 (first approach using all features), which amounted to 0.321 in terms of MiAP and 0.2547 in terms of GMiAP (7th out of 18 compteting teams), and Run 5 which led to an F-ex score of 0.495 (6th out of 18 teams). 1