Riemannian Set-level Common-Near-Neighbor Analysis for Multiple-shot Person Re-identification

Multiple-shot person re-identification deals with the problem to build the correspondence between the images of the same person appearing at different sites captured by onsite deployed cameras. The difficulty stems from large within-class but small betweenclass variations caused by the change of person appearance and environment. Traditional methods on feature/signature design and/or distance/dissimilarity exploration have been largely investigated, resulting in a quick slowing down of the performance improvement. This paper proposes a novel solution called “Riemannian Set-level Common-Near-Neighbor Analysis” by absorbing the essence of two distinctive and effective state-of-the-art models. More concretely, it generates the discriminative covariance-based representation for each set of images following the Mean Riemannian Covariance Grid approach, while at the same time creatively realizes the set-level neighborhood information based ranking inheriting the key idea of samplelevel Common-Near-Neighbor Analysis. Experiments have been conducted on widely-used benchmark datasets, showing significant performance improvement over state-of-the-art methods.