Unsupervised relative attribute extraction

The quality of supervision in the attribute learning step for image classification is directly proportional to the experience of subjects, and it is a labour-intensive job. Additionally, within and between class variance in the image data make it even insufficient to use attributes categorically. In this paper, a new approach is proposed for unsupervised extraction of relative attributes in image classification to overcome the aforementioned restraints at scalable, low cost and moderate accuracy. The proposed approach has been compared to other attribute based methods available in the literature using the same data sets and experimental conditions; and satisfactory results are achieved.