A Research On Feature Extraction Method For Deep Learning Based Person Re-identification

Person re-identification(re-ID)aims to retrieve person across different cameras.Generally,there are three steps in person re-identification: feature extraction,similarity measure and rank.Among them,getting person feature with high quality is important to improving re-ID accuracy.However,due to the movement of person,detection error of person detection algorithm,background differences across cameras and so on,there are several different kinds of misalignment in images.These kinds of misalignment can cause bias in extracted pedestrians' feature and decrease the accuracy of re-ID.In our paper,we discuss the characters in pedestrians' images and divide the misalignment among them into three types: pose change,orientation change and misdetection.With respect to the three types of misalignment,we propose three deep learning based solutions,separately.Moreover,based on three solutions proposed above,we propose a bias-free feature extraction network.Our contributions are listed as follows:With respect to pose change under different cameras,we propose an attention based method to extract discriminative local feature.Visualization and tests on several datasets prove the effectiveness of our method.With respect to the orientation change of pedestrians,we propose orientation sensitive feature extraction model.That can remove the interference from different orientations and generate fine-grained pedestrians' feature,which is adapt to the change of orientation.With respect to misdetection in re-ID data,we design misdetection correction model.By using spatial transformation to align the images,the model effectively remove the interference introduced by misdetection.Because different kinds of misalignment are existing simultaneous,we merge three methods above into one unified structure and propose a feature extraction network that can remove most of the misalignment.Visualization and tests on several large scale datasets prove that our method can remove different kinds of misalignment interference and improve accuracy of re-ID.