Study On Mask-Guided Region Attention Network For Person Re-Identification

Person re-identification(ReID)is an important and practical task which identifies pedestrians across non-overlapping surveillance cameras based on their visual features.In general,Re ID is an extremely challenging task due to complex background clutters,large pose variations and severe occlusions.To improve its performance,a robust and discriminative feature extraction methodology is particularly crucial.Recently,the feature alignment technique driven by human pose estimation,that is,matching two person images with their corresponding parts,increases the effectiveness of Re ID to a certain extent.However,there are still a few problems among these methods such as imprecise handcrafted segmentation of body parts,severe self-occlusions and so on.This thesis focuses on the study and improvements of these problems with its detailed work summarized as follows:1.Due to the large pose variations,the features will be misaligned,and meanwhile,the complex background clutters and severe occlusions will easily lead to the extracted features containing noise,so this thesis presents a novel network model accordingly,which is called Mask-Guided Region Attention Network(MGRAN).MGRAN consists of two major components:Mask-guided Region Attention module(MRA)and Multi-feature Alignment module(MA).MRA aims to generate spatial attention masks to mask out the background clutters and occlusions.Meanwhile,the generated masks are utilized for region-level feature alignment in the MA module.This thesis evaluates the proposed method on three public datasets,including Market-1501,Duke MTMC-re ID and CUHK03.Extensive experiments show the effectiveness of this model.2.Based on the overall thinking of MGRAN,this thesis further improves the recognition accuracy and recognition speed of the model.Because the MRA module in MGRAN is based on the object detection models under the anchor mechanism,this thesis improves the recognition accuracy of MGRAN by using the better object detection model YOLOv4,and meanwhile,this thesis makes MGRAN model obtain the object detection results with higher localization accuracy and the pedestrian image retrieval performance with higher speed respectively by using the more appropriate Non-Maximum Suppression algorithm(NMS)Softer-NMS in object detection and the kd tree data structure.These improvements make MGRAN further optimized and improved in recognition accuracy and speed.On the same three datasets as above,extensive experiments show the effectiveness of the above improvement schemes.