Deep Models For Scale-adaptive Person Re-identification

Surveillance cameras are ubiquitous in modern cities,generating a large amount of video data every day.The increased data requires efficient machine processing capabilities.This paper mainly focuses on the analysis of pedestrians in surveillance video,that is,Person Re-identification(Re-ID).The task of Re-ID is to match person in non-overlapping camera views.In realistic scenarios,the captured pedestrian images may have inconsistent dimensions due to the different camera positions,and the multi-scale pedestrians seriously degrade recognition performance.However,most existing Re-ID methods lay emphasis on matching normal-scale high-resolution person images,and public datasets lack samples for low-resolution pedestrians.To address this problem,this article focuses on the research of scale-adaptive person re-identification,the main points are depicted as follows:1.Low-resolution pedestrian images have much less semantic meaningful information than normal-scale pedestrians.Unsupervised Domain Adaptation(UDA)has a good effect on processing inter-domain distribution migration,so this article attempts to use UDA for scale-adaptive Re-ID research.During training,the UDA model has labeled HR source images and unlabeled multi-scale target images.During test,the LR query images will match the normal-scale gallery images.Experimental results show that the UDA method can improve the recognition performance to a certain extent while still needs to be improved in practical application.2.The image super-resolution algorithm can reconstruct low-resolution images into high-resolution images.This paper proposes a new joint end-to-end learning method for Scale-Adaptive Person Super-Resolution and Re-identification(SASR~2).This method realizes scale-adaptive Re-ID by jointly learning for super-resolution and re-identification without any post-processing process.It can be adaptable to person Re-ID on both multi-scale LR and normal-scale HR datasets,and the super-resolution module effectively improves the model recognition performance.3.In view of the complex feature space nonlinear mapping between low-resolution and high-resolution images,this paper proposes a Semi-Coupled learning model(SCM)to solve the scale-adaptive Re-ID problem.The SCM model consists of two convolutional networks whose inputs are LR images and HR images.Between the two networks,the shared filters learn the mapping of the feature space,and the unshared filters learn the specific resolution.Experiments show that for single scale low-resolution images,the recognition performance of the SASR~2 model is better;for multi-scale images,SCM has comparable performance with the SASR~2 model by a simple semi-coupled learning method,and its implementation is simpler and more suitable for practical applications.