Searching an appropriate template size for multimodal image clustering

It has been shown by researchers that using a multimodality approach can help in identifying better clusters in an image collection. The multimodal image features include low-level image features and available text annotations. This approach helps in identifying inherent relationships among different types of features associated with an image. In our approach, we divide images into small tiles and create visual keywords using a high-dimensional clustering algorithm. These visual keywords act the same as text keywords. One of the challenges of this approach is to identify an appropriate size for visual keywords. In this paper, we report our results in finding a suitable template size that can be used to create tiles for visual keywords. These visual keywords are combined with text keywords to create a multimodal image representation before applying clustering.