Person Re-Idenfitication Based On Deep Conditional Random Field

Person re-identification(Reid)is an active research in the field of computer vision and plays an important role in smart city,intelligent security and public safety.It aims at identifying images or videos of the same person across different cameras.However,challenges still remain due to the diversity of images caused by person’s properties and external environment.Deep learning technique has made breakthrough development in recent years,showing powerful representation learning capability in Reid tasks.Deep learning based approaches have become the mainstream in the Reid community.In essence,person Reid can be regarded as a sorting problem based on similarity learning between a query image and gallery images.However,most deep learning based methods consider correlations over small number samples,ignoring abundant relations of other samples in the dataset,which results in local optimal solution of a Reid network.Graph models have recently been proven effective to address the above issue because of the ability to modeling correlations among data.The core of a graphical model lies in that message passing among graph nodes is exploited to construct the relationships of samples in a training batch,thus achieving more discriminative global decision.From the perspective of modeling global relationship with graph models,this paper integrates conditional random field(CRF)with deep learning technique and conducts an investigation into deep CRF based person re-identification.The main contributions of this work are as follows:(1)Person re-identification by global group-wise similarity learning with CRF.Most existing deep learning Reid approaches employ local similarity strategy,i.e.,exploiting the relationship between a given query and a few gallery samples while ignoring relations among other gallery samples in a training batch.To overcome this limitation,a person Reid algorithm is designed by CRF inference with global group-wise similarity.Specifically,unary potential is built on a siamese network to model local similarity for each query-gallery image pair.The pairwise potential is defined as the formulation related to the unary potential,which is capable of modeling relations of each gallery-gallery pair.Finally,the solution of global similarity is formulated as inference problem by minimizing potential costs of CRF nodes.The mean field inference is conducted to propagate relationships in both unary and pairwise potentials,enabling global group-wise similarity learning in a training batch.Ablation studies indicate the effectiveness of each module in our CRF framework,and the proposed method achieves competitive results on three widely-used person Reid datasets.(2)End-to-end deep CRF model for person re-identification.There are two limitations in above method.First,the definition of pairwise potential is related to the unary potential,which has negative impact on the effectiveness in model learning.The second is that the mean field inference is only conducted in the training stage,not involved in testing.To address these issues,a novel end-to-end deep CRF model is proposed,where CRF potential functions are learned implicitly with deep neural networks and inference for minimization of CRF potential costs is defined as gradient descent based state updating.Since all the modules involved are formulated into standard neural network operations,CRF potentials and global group-wise similarity learning are optimized jointly in an end-to-end manner.In specific,we also use a siamese network to model unary potentials.A bidirectional long short term memory(LSTM)is utilized to construct pairwise potentials over arbitrary node pair.By gradient descent algorithm,the CRF inference is formulated as the iteration of recurrent neural network(RNN).Thanks to the differentiability of each module,the optimization of CRF potential learning and CRF inference are integrated into an end-to-end trainable manner.A series of ablation studies are conducted to validate the contributions of each module.Extensive experiments are conducted on three largescale person Reid datasets,and the results demonstrate the effectiveness of the proposed method by comparison with graph-based approaches as well as several state-of-the-art methods.