Although the performance of person Re-Identification (ReID) has beensignificantly boosted, many challenging issues in real scenarios have not beenfully investigated, e.g., the complex scenes and lighting variations, viewpointand pose changes, and the large number of identities in a camera network. Tofacilitate the research towards conquering those issues, this paper contributesa new dataset called MSMT17 with many important features, e.g., 1) the rawvideos are taken by an 15-camera network deployed in both indoor and outdoorscenes, 2) the videos cover a long period of time and present complex lightingvariations, and 3) it contains currently the largest number of annotatedidentities, i.e., 4,101 identities and 126,441 bounding boxes. We also observethat, domain gap commonly exists between datasets, which essentially causessevere performance drop when training and testing on different datasets. Thisresults in that available training data cannot be effectively leveraged for newtesting domains. To relieve the expensive costs of annotating new trainingsamples, we propose a Person Transfer Generative Adversarial Network (PTGAN) tobridge the domain gap. Comprehensive experiments show that the domain gap couldbe substantially narrowed-down by the PTGAN.
展开▼