Object Lesson: Discovering and Learning to Recognize Objects

摘要

Statistical machine learning has revolutionized computer vision. Systems trained on large quantities of empirical data can achieve levels of robustness that far exceed their hand-crafted competitors. But this robustness is in a sense 'shallow' since a shift in context to a situation not explored in the training data can completely destroy it. This is not an intrinsic feature of the machine learning approach, but rather of the rigid separation of the powerfully adaptive training phase from the final cast-in-stone system. An alternative this work explores is to build 'deep' systems that contain not only the trained-up perceptual modules, but the tools used to train them, and the resources necessary to acquire appropriate training data. Thus, if a situation occurs that was not explored in training, the system can go right ahead and explore it. This is exemplified through an object recognition system that is tightly coupled with an 'active segmentation' behavior that can discover the boundaries of objects by making them move.