Fine-grained Collaborative K-Means Clustering

We present a collaborative clustering algorithm of fine-grained images, where the subtle inter-class differences are represented using collaborative filters. Cluster centers are represented as optimal weighted collaboration of data points, and the optimal weight matrix is analytically obtained. This may be viewed as a generalization of the K-means clustering algorithm, where these weights would be unity. We also introduce a matrix inversion scheme that allows us to scale up collaborative representations by orders of magnitude and this allows us to apply the scheme when the number of data points is large. Collaborative clustering outperforms K-means and a few of its popular variants K-medians, K-modes and K-medoids. It also outperforms DBSCAN and its recent variation DSets-DBSCAN. We have chosen species recognition (bird and butterflies) as a representative fine-grained image categorization problem, though the proposed algorithm is a general approach applicable to other similar tasks. We also introduce a new benchmark fine-grained image dataset, that of Indian endemic butterfly species (Titli.vl), which is available through the corresponding author.