Person Re-identification Based On Convolutional Neural Networks

Person Re-identification(Person Re-ID)is the process of discriminating whether two person images,which are respectively taken from two cameras with no overlap in field of view,belong to the same person.Person Re-ID is a promising technology which can be applied in intelligent video surveillance of criminal investigation.Person Re-ID can also be used to produce traffic data,which is helpful in the designing and updating of traffic systems and markets' layout.Due to the illumination variance across different cameras,the angle of view variance across different cameras,the person's pose variance across different cameras,and the occlu-sion issue,significant difference usually exists between person images captured in different camera views,even though the two images belong to the same person.With the help of deep learning technology,the state-of-the-art evaluation results on several large scale Person Re-ID dataset has been significantly increased.However,the performance of state-of-the-art Person Re-ID methods is far from satisfying the application demand.In this paper,we design a new type of convolutional neural networks,which processes Person Re-ID by comparing feature maps of two person images.Specifically,the main contributions of this paper are as follows.1.Misalignment between person images and color differences across cameras are two major problems in feature map comparison based Person Re-IED.Due to the two problems,comparing feature maps by calculating elementwise-form Euclidean distance or L2 distance lead to unsatisfactory Person Re-ID results.To solve abovementioned problem,we propose a novel CNN architecture named BraidNet.BraidNet has a kind of specially designed WConv layer,which has two input feature maps in the same size and two output feature maps in the same size.In a BraidNet,feature maps of two input person images are firstly extracted by one subnetworks;then these feature maps are fed into the cascaded WConv structure(the cascaded WConv structure contains ReLU and Pooling layers);the two output feature maps are elementwisely added together to make one feature map;finally,the feature map is fed into another subnetworks to get one matching score,which indicates whether the two input person images belong to the same person.Theoretically,a single WConv layer or a cascaded WConv structure can dig cues which are helpful for the final matching/not match-ing decision,from the misalignments and color differences across cameras situations be-tween the two input person images.2.When the ReLU function is chosen as the activation function of one CNN model,training this model faces the zero gradient problem,which means that the responses on some channel are always non-positive regardless of the inputs,and corresponding weights to cal-culate these responses can not get non-zero gradient to update their value.The fitting capac-ity of one CNN model may be decreased during the training phase due to the zero gradient problem.We propose a simple channel scaling layer to handle the zero gradient problem.In one channel scaling layer,the responses on each input channel are multiplied by a positive definite factor to obtain the responses on corresponding output channel,and these positive definite factors can be learned during the training phase.We qualitatively proved that using the channel scaling layer can help alleviate the zero gradient problem.What's more,we recognized some correlations between the channel scaling layer and some CNN model prun-ing methods,which demonstrates that the channel scaling layer can learn the importance of each input channel during the training phase.3.The BraidNet is a binary classification model,and faces the problem of imbalanced training data(imbalanced amounts of positive samples and negative samples)in the training phase.To solve this problem,the sample rate learning strategy is proposed.With sample rate learning strategy,the percentage of positive samples and negative samples in each batch can be adaptively adjusted during the training phase,so as to avoid the puzzle of choosing the optimal hand-crafted setting of the percentage of positive samples and negative samples in each batch.On CUHK03-Detected?CUHK03-Labeled?CUHK01?Market-1501?DukeMTMC-reID and DukeMTMC4ReID datasets,we verify the effectiveness of abovementioned three contributions by a set of contrast experiments and visualization of features.Finally,we dis-cuss the characteristics of BraidNet and provide an outlook of BraidNet's application.