Learning Rotational Invariant Dictionary for Sparse Coding based Key-point Detection

Recent research has shown the effectiveness of a Scale and Rotational Invariant Sparse Coding based Key-point Detector. However, this detector utilizes a dictionary generated by combining a carefully selected seed dictionary with multiple rotated versions of its atoms. This manual generation process is time-demanding, thus a novel autoencoder structure for automating it is proposed. Specifically, the weights between input and hidden layers of the structure are designed to embed a rotational invariant dictionary. Indeed, it is a duplet structure constrained by a novel loss function such that rotated inputs correspond to circularly shifted versions of hidden layer responses. To our knowledge, an autoencoder embedded in its weight a rotational invariant dictionary is new and the proposed loss function is the first to fulfill this requirement. Experimental results show that the learned dictionary preserves the original detector's performance, helps achieve state-of-the-art key-point detection performance while substantially eliminating the laborious work.