Learning Object Categories with Generic Knowledge

The dream of computer vision is a machine capable of interpreting images of complex scenes. Central to this goal is the ability to recognize objects as belonging to classes and to localize them in the images. In the traditional paradigm, each new class is learned starting from scratch, typically from training images where the location of objects is manually annotated (fully supervised setting). In this work instead, knowledge generic over classes is first learned during a meta-training stage from images of diverse classes with given object locations. This generic knowledge is then used to support the learning of any new class without location annotation (weakly supervised setting). Generic knowledge makes weakly supervised learning easier by providing a strong basis: during meta-training the system can learn about localizing objects in general. This strategy enables learning from challenging images containing extensive clutter and large scale and appearance variations between object instances, such as the PASCAL VOC 2007. In turn, this opens the door to learning a large number of classes with little manual labelling effort.