Dispersion based clustering for unsupervised person re-identification
Abstract
The cumbersome acquisition of large-scale annotations for person re-identification task makes its deployment difficult in real-world scenarios. It is necessary to teach models to learn without explicit supervision. This paper proposes a simple but effective clustering approach for unsupervised person re-identification. We explore a basic concept in statistics, namely dispersion, to achieve a robust clustering criterion. Dispersion reflects the compactness of a cluster when assessed within and reveals the separation when measured at the inter-cluster level. Based on this insight, we propose a Dispersion based Clustering (DBC) approach which performs better at discovering the underlying data patterns. The approach can automatically prioritize standalone data points and prevents poor clustering. Our extensive experimental results demonstrate that the proposed methodology outperforms the state-of-the-art unsupervised methods on person re-identification.