Personal re-identification using rank-based manifold ranking

In this paper, a rank-based manifold ranking (MR) algorithm is proposed for personal re-identification. In general cases, L1 norm, L2 norm, or cos production metrics are frequently adopted for distance or similarity calculation with a heat kernel function. However, outliers in the distance-based scheme always impact the identification results even though the number of outliers is few. A rank-based weighting scheme is adopted in the MR algorithm instead of the distance-based metric. A benchmark dataset of video surveillance is evaluated. The experimental results are given to show the feasibility of the proposed method.