Résumé
This paper discusses the results of the LifeCLEF 2014 multimedia identification challenges with regards to the requirements of real-world ecological surveillance systems. In particular, we study the identification performances of the evaluated systems as a function of the ordinariness or rarity of the species in the dataset. This allows us to assess the ability of the underlying methods to be robust to heavily tailed distributions such as the ones encountered in real-world collections of life observations. Results show that all methods are more or less affected by the long-tail curse but that the best methods making use of classifiers with good discrimi- nation capacities do resist the phenomenon pretty well.