Study On Metric Learning Based Person Re-identification

Person re-identification aims to associate the identity of individuals observed from different cameras with non-overlapping fields-of-view.It serves as a critical step in applications like multi-camera tracking,multi-view action analysis,pedestrian retrieval,and so on.Due to the low quality of surveillance video and uncontrolled environment,the biometrics such as face,finger print,and gait are usually unreliable or infeasible.As a result,person re-identification has to rely upon the visual appearance which may change significantly in different camera views due to large variations in illumination,body pose,viewpoint and occlusion.Therefore,it is rather difficult to identify pedestrians by their visual appearances.In current person re-identification literature,the metric learning based methods are more prevailing with impressive performance.From the metric learning viewpoint,this paper discusses some important problems such as the Small Sample Size(SSS)problem,the impact of imbalanced samples,and so on.And then profound research has been made.The main contents of this paper include the following four aspects.(1)For the problem that the metric may be biased by unbalanced samples involved in most person re-identification datasets,a Large Margin Relative Distance metric Learning(LMRDL)algorithm is proposed.By adopting the strategy of learning the distance metric from triplets under relative comparison constraint,the problem of imbalanced samples can be addressed efficiently in LMRDL.The proposed LMRDL is formulated as a convex logistic metric learning problem which guarantees the existence of global minimum.For optimization,the Accelerated Proximal Gradient(APG)approach is employed to find the optimal solution.Experimental results show that LMRDL can improve the re-identification performance effectively when compared with the models that learn from pairwise constraints.(2)To exploit the discriminative information of both distance and bilinear similarity provided by the training data,a Generalized Similarity Metric Learning(GSML)algorithm is proposed.In GSML,the Mahalanobis distance function and the bilinear similarity function are combined to promote the complementary advantages of jointly learning two different metrics.Experimental results demonstrate that much better re-identification performance can be obtained than learning the metrics separately.The proposed GSML also outperforms the state-of-the-art approaches and LMRDL significantly.(3)In order to tackle the Small Sample Size(SSS)problem suffered by most distance metric learning methods for person re-identification,we propose an algorithm called Kernel Marginal Nullspace Learning(KMNL).The KMNL learns a discriminative nullspace to avoid computing the inverse of a singular within-class scatter matrix induced by SSS problem.In the learned nullspace,multiple images of the same pedestrian are collapsed to a single point.Therefore,it is optimal to map pedestrian images to the nullspace for cross-view matching if the between-class distances are non-zero.With rigorous theoretical analysis,we first derive the closed-form solution of the nullspace in the linear case,and then extend it to the non-linear scenario with kernel trick.Experimental results show that KMNL not only addresses the SSS problem effectively,but also can achieve impressive re-identification performance with low training time cost.(4)A novel person re-identification framework named Iterative Multiple Kernel Metric Learning(IMKML)is further developed based on the KMNL algorithm.The IMKML can simulate the online collection of training samples and decision gathering of multiple experts in manual re-identification procedure.It exploits the discriminative information from not only the training set but also the successfully identified test pairs.Specifically,there are two main modules in the IMKML framework.The first module learns multiple kernel nullspaces via KMNL,and the other is to pick out the most probable positive test pairs to augment the training set.By iteratively alternating the two modules,much more samples will be involved for training and significant performance improvement can be gained.This paper presents a series of solutions for the problems in current metric learning based person re-identification,and extensive experiments are carried out to demonstrate their effectiveness.The achievements will further promote the applications of person re-identification in video surveillance.