Reconocimiento de ubicación con lapsos de tiempo para la anotación automática de videos antiguos

Abstract

Los algoritmos basados en el aprendizaje han tenido mucho éxito recientemente en un amplio rango de problemas de visión artificial. El uso extendido de dispositivos móviles dotados de cámaras, sensores GPS y conexión a internet, ha generado interés por el problema de reconocimiento de ubicación visual, ya que es visto como una herramienta valiosa en aplicaciones de turismo y de navegación. In visual recognition systems it is often the case that the conditions in which a model is developed differ from those in which the model is deployed. The performance of a system that has been trained on label data from a source domain is severely affected when tested on samples drawn from a different target domain distribution. This issue imposes a limit on the real-world applications of a system, as expensive manually annotated data is needed for every new domain. Domain adaptation aims to handle this mismatch between distributions, and to develop models that will perform well on the test data in the new domain. In this work the explored domain shift is caused by a temporal gap between ancient photographs and frames of movies and modern images, and the task considered is that of location recognition. The Geodesic Flow Kernel and the Subspace Alignment domain adaptation methods are evaluated on two datasets. The first one consists of historical and modern images depicting 25 locations of cities across the world, and the second one consists of frames of 1945 Italian movie Rome, Open City and modern geotagged images that cover the same set of locations downloaded from Google Street View. We find that the classification performance improves when domain adaptation, together with state-of-the-art features, are used on the task of location recognition with a temporal gap between training and test sets.