Learning Image Attributes using the Indian Buffet Process

In the domain of object recognition and image classification, a recent trend is to use image properties or attributes to represent the images. Most of the proposed models in the past require that the number of attributes and attribute semantics be specified in advance. In this paper, we propose a generative model for image attributes that combine attribute-based vision models and feature-based nonparamatric models. We learn the model using Gibbs sampling. Qualitatively, we demonstrate the learned attributes of images in three categories. Quantitatively, we show that our model outperforms simple baseline methods in image retrieval and transfer learning tasks.