Large-Scale Incremental Visual Learning Using Rich Feature Hierarchies
Svetlana Lazebnik, University of Illinois at Urbana-Champaign
Usage Details
Svetlana Lazebnik, Juan Caicedo Rueda, Liwei WangBuilding accurate and generic vision systems requires features that can discriminate contents in a wide variety of images. Deep convolutional neural networks (CNN) are powerful tools to learn these features automatically from available data. We address the problem of expanding the capacity of a CNN using large-scale image collections from the Web. Our goal is to design strategies to progressively learn more accurate models over time as more data becomes increasingly available. This contrasts with the common practice of training vision models with a static dataset, which means that the resulting model needs to be retrained when the dataset changes.