Discovering Multi-relational Latent Attributes by Visual Similarity Networks

The key problems in visual object classification are: learning discriminative feature to distinguish between two or more visually similar categories (e.g. dogs and cats), modeling the variation of visual appearance within instances of the same class (e.g. Dalmatian and Chihuahua in the same category of dogs), and tolerate imaging distortion (3D pose). These account to within and between class variance in machine learning terminology, but in recent works these additional pieces of information, latent dependency, have been shown to be beneficial for the learning process. Latent attribute space was recently proposed and verified to capture the latent dependent correlation between classes. Attributes can be annotated manually, but more attempting is to extract them in an unsupervised manner. Clustering is one of the popular unsupervised approaches, and the recent literature introduces similarity measures that help to discover visual attributes by clustering. However, the latent attribute structure in real life is multi-relational, e.g. two different sport cars in different poses vs. a sport car and a family car in the same pose - what attribute can dominate similarity? Instead of clustering, a network (graph) containing multiple connections is a natural way to represent such multi-relational attributes between images. In the light of this, we introduce an unsupervised framework for network construction based on pairwise visual similarities and experimentally demonstrate that the constructed network can be used to automatically discover multiple discrete (e.g. sub-classes) and continuous (pose change) latent attributes. Illustrative examples with publicly benchmarking datasets can verify the effectiveness of capturing multi- relation between images in the unsupervised style by our proposed network.