Improving object classification using semantic attributes

This paper shows how semantic attributes can be used to improve object classification. The semantic attributes used fall into five groups: scene (e.g. ‘road’), colour (e.g. ‘green’), part (e.g. ‘face’), shape (e.g. ‘box’), and material (e.g. ‘wood’). We train a set of classifiers for individual semantic attributes, and use them to make predictions on new images (Figure 1). We can then use the scores from the set of classifiers as a low-dimensional image representation. The object classification performance of the semantic attribute features alone is close to that of a much higher-dimensional bag-of-words image representation, while using the semantic attributes together with the bag-of-words features consistently improves performance.