An improved pedestrian re-identification algorithm based on unsupervised domain adaptation
In recent years,the application of pedestrian re-identification has become more and more widespread.In order to better solve the problem of large differences in cross-domain adaptive recognition due to resolution,illumination,etc.and the diffi-culty of labeling a large number of data sets,a method based on The unsupervised adaptive pedestrian re-identification network(PU-Net)for pseudo-label generation uses the DBSCAN clustering algorithm to generate pseudo-labels,uses an improved residual network to extract features,and adds a channel attention mechanism to improve feature extraction capabilities.Experimental re-sults show that on the DukeMTMC-ReID,Market-1501 and MSMT17 data sets,mAP increased by 2.8,0.7,and 0.3 percentage point respectively;Rank-1 increased by 2.7,0.3,and 0.9 percentage point respectively.Compared with most models,this method achieves significant improvements and is close to the results of fully supervised training.
person re-identificationunsupervised adaptationpseudo-labelhannel attention
万方数据
