Texture Characterization with Semantic Attributes: Database and Algorithm

Categorization is an important capability of service robots for understanding their environment. While categorization may base on features of different modalities, such as 3D shape, size, or 2D interest points, this paper focuses on categorizing textures. In contrast to previous approaches that base texture classification on rather abstract numerical features, this work introduces a set of 17 properties that can easily be interpreted by a human establishing new options on human robot communication and learning. A new database was recorded to account for the service robotics context yielding almost 1500 images divided into 57 classes of textures and objects. The dataset is fully annotated with the 17 attributes which represent color and structural properties measured on a continuous scale. The evaluation on this database compares five methods for attribute learning, texture classification, and zero-shot learning with promising results.