Robust attribute-based visual recognition using discriminative latent representation

Abstract

Recent work in visual recognition have addressed attribute-based classification. However, semantic attributes that are designed and labeled by humans generally contain some noise, and have weak learnability for classifiers and discrimination between categories. As a fine supplement to semantic attribute, data-driven attribute learned from training data suffers from the ineffectiveness in novel category classification with no or few samples. In this paper, we introduce the Discriminative Latent Attribute (DLA) as a mid-level representation, which has connection with both visual low-level feature and semantic attribute through matrix factorization. Furthermore, we propose a novel unified formulation to efficiently train category-DLA matrix and attribute classifiers together, which makes DLA more learnable and more discriminative between categories. Our experiments show the effectiveness and robustness of our approach which outperforms the state-of-the-art approach in zero-shot learning task.