Open World Recognition

Abstract: I mage classification module has difficulty in identifing a test image from untrained categories. I think this is because neural network cannot transfer untrained categories into ranges disjoint from ranges of the trained categories.  Image recognition is very hot topic in the computer vision field. As a fundumental module of the image recognition system, image classification has been studied for decades and remains an active task. Since the AlexNet in 2012, deep learning becomes the main methodology in doing image classification and it indeed achieves impressive preformance increases in many datasets like cifar10 or ImageNet. However, those dataset setups are under sort of closed world assumption. That is, train set and test set have exactly the same categories with similar sample distribution. This closed world assumption, although reasonable for computer vision scientists/engineers, is difficult for others to understand and in fact is invalid in real world. In real ...