Shape classification using invariant features and contextual information in the bag-of-words model

• A shape classifier tolerant to scale, rotation and viewpoint changes is proposed.
• The inclusion of the histogram of bi-grams improves the bag-of-words model.
• Codebook selection and feature learning improve the classification accuracy.
• We report best results on the animal shapes dataset using the proposed method.

In this paper, we describe a classification framework for binary shapes that have scale, rotation and strong viewpoint variations. To this end, we develop several novel techniques. First, we employ the spectral magnitude of log-polar transform as a local feature in the bag-of-words model. Second, we incorporate contextual information in the bag-of-words model using a novel method to extract bi-grams from the spatial co-occurrence matrix. Third, a novel metric termed ‘weighted gain ratio’ is proposed to select a suitable codebook size in the bag-of-words model. The proposed metric is generic, and hence it can be used for any clustering quality evaluation task. Fourth, a joint learning framework is proposed to learn features in a data-driven manner, and thus avoid manual fine-tuning of the model parameters. We test our shape classification system on the animal shapes dataset and significantly outperform state-of-the-art methods in the literature.