Unsupervised Person-reidentification Based on Cross-camera Clustering Comparison
Unsupervised pedestrian re-identification has received increasing attention in industry due to its high label cost.However,most unsupervised REID works generate pseudo labels by clustering and comparing the similarity between features with-out considering the distribution differences between cameras.This can lead to a decrease in the accuracy of inter camera labels,meaning that the trained model cannot guarantee good deployment on cameras that have never been seen before.This problem great-ly limits the application of re-identification algorithms.In order to solve this problem,this paper proposes a camera based clustering contrast between cameras,which uses the instance batch processing normalization module(CIBN)based on cameras to normalize pedestrian images collected by all cameras into subspaces,narrow the distribution gap between cameras,and adjust the proportion of adjustable parameters according to different experimental data sets to combine IN and BN optimization,greatly improving the gen-eralization ability of the model.Secondly,combined with clustering comparison,the pre calculated instance feature vectors are stored in memory,and pseudo labels are assigned to them using clustering algorithms.Then,a comparison loss function is used to compare the query instances with the clusters,and finally the dictionary is updated.Compared with unsupervised REID Pipeline us-ing ordinary BN algorithm,this method achieves significant improvement.On the Market,Duke,and MSMT17 datasets,mAP has improved by 7.4%,13.7%,and 8%respectively compared to state-of-the-art pure unsupervised methods.
person re-identificationunsupervisedclustering comparisonpseudo label
万方数据
维普
