Optimisation de l’´echelle d’observation pour l’annotation d’images

mathis cordier, Pejman Rasti, Cindy Torres, David Marc ROUSSEAU

Abstract: Data annotation is currently a bottleneck in the development of Deep Learning. Indeed, we have very efficient models and sufficiently large computing capacities to train these models. On the other hand, it is complicated to obtain sufficiently large annotated databases for them, especially when the annotation needs to be done by a domain expert. In the context of detecting disease symptoms on plants, we need these annotations to be done by specialists and we set up annotation campaigns with them. We are then looking for a way to make these annotation campaigns more efficient, by reducing the annotation time for the specialists and thus reducing their fatigue. We show in this paper the existence of an optimal observation scale for these annotations. Our empirical approach is based on the statistical analysis of eye-tracker signals. It opens the way to more methodological questions.