Closing the gap between weakly and fully supervised methods

Опубликовано 17 августа 2016, 3:48

Learning visual classes is traditionally done in a fully supervised setting, where the location of training instances are manually annotated by bounding-boxes or pixelwise segmentations. After learning a model of a class from this data, it can be used to recognise and localise novel instances in test images. On the other hand, weakly supervised techniques try to learn such models from training images labeled only by the presence or absence of the class, without location annotation. While these techniques can substantially reduce the manual effort necessary to learn a class, fully supervised methods typically deliver models that perform considerably better on test data. In this talk I will present recent advances towards closing this performance gap for the two tasks of action recognition and semantic segmentation. These are steps towards the goal of devising weakly supervised techniques that deliver models of the same quality as fully supervised ones.