Recognition of Complex Objects by Hierarchical Attributes

Most of classification models are trained under the guidance of dataset labeled by category. It is likely to fail to extract some important features due to limited knowledge. That is, as to the complex objects, it is hard to learn the complete feature representations by this means. So a prior knowledge is required to guide the learning of the model. With respect to this problem, this paper proposes a method of classifying objects assisted by hierarchical attributes. The attribute refers to the high-level description of objects, which can be semantic (e.g. small mammals) or discriminative (e.g. having four legs). This method extracts the category-level features and attribute-level features from an object, and conducts classification after fusing these features. Among them, the attribute-level features provide extra evidences for classification. In the fine classification task on CIFAR−100, the proposed method can obtain higher accuracy than general methods. Additionally, this method can be extended to image retrieval via dividing the indexing space by attribute-level labels.