Domain Transfer Learning For Person Re-identification

Current person re-identification(re-ID)algorithms focus on supervised methods using deep neural networks(DNNs).While models learned with labeled data have demonstrated human-level performance in the laboratory setting,which is prohibitive to deploy such models in the real-world scenario.One crucial cause of this problem is the distributional variation between the training and test data.It is possible to mitigate this variation by collecting i.i.d training data for each test scenario,but that requires exceptional monetary and time costs.As such,researchers have started applying domain transferring techniques to address the real-world deployment issue of person re-ID.This work makes contributions to two fundamental problems in the domain transferring area for person re-ID:unsupervised domain adaptation(UDA)and domain generalization(DG).The contributions of this work are summarized as follows.(1)Dynamic task-oriented feature disentanglement for UDA.This work proposes a novel concept,"task-oriented" and a dynamic task-oriented disentangling network(DTDN).DTDN decouples features into task-relevant and task-irrelevant components.Task-relevant parts identify both task-effectiveness and domain-irrelevance,while task-irrelevant parts identify at most one of them.DTDN can effectively facilitate UDA performance in person re-ID.Empirical benchmarks on Market-1501 and DukeMTMC-ReID demonstrate 8.9%and 6.7%improvements in Rank-1 accuracy.(2)Dual distribution alignment network for DG.This work identifies two fundamental challenges in DG for person re-ID:domain invariance and identity similarity.To address these challenges,it proposes two novel constraints,namely the peripheral-center domain adversarial feature learning and cross-domain identity similarity enhancement.A dual-distribution alignment network(DDAN)is proposed under such constraints to learn generalizable and domain-invariant features.DDAN can effectively reduce the variation among multiple domains,allowing for model deployment without data re-collection or model re-training.Empirical benchmarks demonstrate 3.3%improvements in Rank-1 accuracy over state-of-the-art approaches.