View Confusion And Compact Feature Learning Based Person Re-identification

Person re-identification refers to the searching and matching for a specific pedestrian in a cross-camera environment.At present,because of the development of deep learning and computer vision,the performance of person re-identification is also greatly improved.However,there are still many challenges in person re-identification.We propose new person re-identification algorithms based on transfer learning,hash learning and attention mechanism.The main works and contributions of this paper are as follows:(1)A view confusion feature learning algorithm is proposed for person re-identification.The performances of person re-identification tasks can be seriously degraded because of variations caused by view changes.In recent years,there are many methods focusing on how to solve cross view challenges which can be roughly divided into two categories: 1)view-generic models 2)view-specific models.However,methods of the first category are not robust enough for different kinds of view-invariants while methods of the other category can't generalize well in real-world applications.In this paper,we combine this two kinds of methods,aiming to learn view-invariant features with the help of view information.We proposed an end-to-end trainable framework,called View Confusion Feature Learning(VCFL),to learn view-invariant features by getting rid of view specific information.Experiments on three benchmark datasets including Market-1501,CUHK03,and Duke-MTMC prove the superiority of our method over state-of-the-art approaches.(2)A view guided attention mechanism algorithm is proposed for person re-identificationMost methods pay attention on the design of attention mechanism,and only a few methods restrict the attention part of the attention mechanism.In order to learn view-invariant features and make the attention mechanism better improve the performance,we try to guide the attention mechanism with view information,so as to pay more attention on the view-invariant parts in the feature map.Further,we can get view-invariant features through the idea of multi-adversarial learning.We conduct experiments on different attention mechanisms,proving the effectiveness of view guided attention mechanism and multi-adversarial learning.(3)A compact feature learning algorithm is proposed for person re-identification.Person Re-ID methods have been much improved in terms of accuracy,however,matching efficiency is rarely considered regarding the rapidly increasing data,which is therefore far from the real-world applications.Thus,we propose deep hashing to enable re-identification more scalable for large-scale gallery set,however,useful information is lost or distorted due to the straightforward hard hash function.In this paper,we propose an end-to-end trainable framework,called Content Inherited Compact Feature Learning(CICFL).And we mainly focus on how to learn discriminate compact features with useful information inherited and negative effects caused by hard hashing function filtered,such that both accuracy and efficiency are promoted beyond a balance.Experiments on three benchmark datasets including Market1501,Cuhk03,and Duke-MTMC show the superiority of our method over state-of-the-arts without losing accuracy.