Non-negative Dual Graph Regularized Sparse Ranking for Multi-shot Person Re-identification
Person re-identification(Re-ID)has recently attracted enthusiastic attention due to its potential applications in social security and smart city surveillance.The promising achievement of sparse coding in image based recognition gives rise to a number of development on Re-ID especially with limited samples.However,most of existing sparse ranking based Re-ID methods lack of considering the geometric structure on the data.In this paper,we design a non-negative dual graph regularized sparse ranking method for multi-shot person Re-ID.First,we enforce a global graph regularizer into the sparse ranking model to encourage the probe images from the same person generating similar coefficients.Second,we enforce additional local graph regularizer to encourage the gallery images of the same person making similar contributions to the reconstruction.At last,we impose the non-negative constraint to ensure the meaningful interpretation of the coefficients.Based on these three cues,we design a unified sparse ranking framework for multi-shot Re-ID,which aims to simultaneously capture the meaningful geometric structures within both probe and gallery images.Finally,we provide an iterative optimization algorithm by Accelerated Proximal Gradient(APG)to learn the reconstruction coefficients.The ranking results of a certain probe against given gallery are obtained by accumulating the redistributed reconstruction coefficients.Extensive experiments on three benchmark datasets,i-LIDS,CAVIARA4REID and MARS with both hand-crafted and deep features yield impressive performance in multishot Re-ID.
Person re-identification Sparse ranking Dual graph regularization Non-negativity
Aihua Zheng Hongchao Li Bo Jiang Chenglong Li Jin Tang Bin Luo
School of Computer Science and Technology,Anhui University,Hefei,China
国际会议
广州
英文
108-120
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）