Research On Cross-Domain Person Re-identification By Image-to-Image Style Translation And Camera-Attribute

The person re-identification(Re-ID)targets at matching images of people under different cameras,and it can be applied to intelligent security,smart city and other fields.With the rapid development of deep learning and convolutional neural network,the person Re-ID model,which is trained with supervision on data collected from a certrain scenario and directly applied in this scenario(i.e.the source domain),has been greatly improved.However,it is still a big challenge to transfer the model to another different scenario(i.e.target domain).This is because the existence of inter-domain difference(completely different persons and background of distinct images collected on two domains)and intra-domain bias(camera view and person pose changes in the target domain).Meantime,in real world applications,there will be multiple test scenarios(i.e.multiple target domains).The cost of training a cross domain model for each target domain is very high.So it is very pratical and challenging to realize a universal crossdomain model that can be used in multiple domains.In this paper,based on deep learning,the model and algorithm of cross domain person re-identification are studied,and we propose two models.The details are as follows:As traditional cross single domain person Re-ID methods directly train cross domain person Re-ID models by transfer learning,it is difficult to accurately bridge inter-domain distribution difference and intra-domain bias for loss functions.To faciliate with this,this paper proposes a novel person Re-ID model based on image-to-image style translation,constrained by a multi-loss function.The whole model is composed of two parts.The first part is the image-to-image style translation model,which aims to learn the distribution of different cameras in source domain and target domain,supplement the absent labels of target domain,expand the data of the target domain,and reduce the impacts caused by inter-domain difference and intra-domain bias.The other part is the person Re-ID model based on the multi-loss function,which aims to learn the feature distribution of persons in the target domain and narrow down the feature distance of the same person.Extensive experiments are conducted on three widely employed benchmarks,including MSMT17,Market-1501 and Duke MTMC-re ID,and experimental results demonstrate that the proposed method can achieve a competitive performance against other state-of-the-art unsupervised cross single domain Re-ID approaches.In terms of cross multi-target domains person Re-ID model in practical applications,there are distribution differences among multi-target domain and it is difficult for a single Re-ID model to learn multi-domain distribution at the same time.To solve with this,on the basis of the across single target domain model,this paper further proposes a universal Re-ID model based on distribution consistency and camera attributes.The model learns the inter-domain distribution of the source domain and multi-target domains at the same time,extracts the camera attributes to distinguish the intra-domain distribution and interdomain distribution,and realizes a universal model for multi-target domains.Extensive experiments are executed on Market-1501,Duke MTMC-re ID and CUHK03,and our model has achieved a very competitive Re-ID accuracy in multiple domains against numerous state-of-the-art methods.