Research On Reranking Approaches For Image-based Person Reidentification

Person reidentification(ReID)refers to the task of verifying the identity of a pedestrian observed from nonoverlapping views in a surveillance camera network.It has recently been validated that reranking can achieve remarkable performance improvements in person ReID systems.However,current rerank-ing approaches either require feedback from users or suffer from burdensome computational costs.In the way of traditional unsupervised learning and deep CNN-based supervised learning,we improve the performance of Reranking Approaches for Image-based Person Reidentification in four perspectives.1.We propose an inverse reranking algorithm by exploring the local neigh-borhood information.In this paper,the concept of k inverse nearest neighbor(k-INN)is deduced as the basic inverse reranking algorithm,and its perfor-mance is evaluated in the person Re-ID task.Based on it,a simple yet effective inverse re-ranking algorithm is proposed,termed inverse density-adaptive k-ernel based reranking(inv-DAKR),which is formulated by a smooth kernel function with a density-adaptive parameter.Experiments on six benchmark data sets confirm that our proposals are effective and efficient.2.We propose a bidirectional reranking algorithm by combining context information from two direction.In this paper,by integrating k inverse nearest neighbor(k-INN)with k nearest neighbor(k-NN),the concept of k reciprocal nearest neighbor(k-RNN)is deduced as the basic bidirectional reranking al-gorithm,and its performance is assessed as one of the baseline.From k-RNN,a simple yet effective bidirectional reranking algorithm is proposed,termed bidirectional density-adaptive kernel based reranking(bi-DAKR),in which the local density information in the vicinity of samples is elegantly exploited in the two direction by adopting a smooth kernel function with a density-adaptive pa-rameter.The experimental results demonstrate that our proposals are effective and efficient.3.We propose the effective reranking algorithm by incorporating incorpo-rate the available extra probe samples.In this paper,inv-DAKR and bi-DAKR evolve into inv-DAKR+and bi-DAKR+by using available extra probe sam-ples in an unsupervised manner,and k-INN+and k-RNN+ are also obtained in the same way and regarded as the baseline to show the effectiveness of ex-tra samples.After delicate analysis,we demonstrate that when and why these extra probe samples are able to improve the local neighborhood and thus fur-ther refine the ranking results,and show the meaning and setting method of hyperparameter k.The experimental results demonstrate that our proposals are effective and efficient.4.We propose an end-to-end deep CNN model which integrates rerank-ing with feature extraction and metric learning.In this paper,we propose a novel approach,called group-shuffling dual random walks with label smooth-ing(GSDRWLS),in which random walks are performed separately on two channels of the FC layer-one for positive verification and one for negative verification-and the binary verification labels are properly modified with an adaptive label smoothing technique before feeding in the verification loss in or-der to train the overall network effectively and to avoid the overfitting problem.Extensive experiments conducted on three large benchmark datasets,including CUHK03,Market-1501 and DukeMTMC,confirm the superior performance of our proposal.