Unknown Object Recognition Using the Manifold Structure of Class Distributions

Recently, most object-recognition techniques have been developed using known classes of objects as inputs. However, in the real world, there is a possibility that "unknown" objects may be used inputs, and conventional methods cannot cope with such cases. In this study, we focus on the unknown object recognition technique, which recognizes an unknown object as an unknown class when it is input, and propose a method to construct a latent space considering the manifold structure of known classes. The basic performance of the proposed method is demonstrated using the CIFAR-10 image dataset, and results of its application to radar image data are also reported.