Cross Domain Person Re-identification Based On Domain Adaption And Incremental Learning

Supervised person re-identification(ReID)has been playing important roles in industry,while the performance of cross-domain ReID consistently lags behind that of supervised scenarios.And considering the limitations of ReID in terms of expensive annotation cost and strict privacy protocol,crossdomain deployment is a necessary and common scenario.To further narrow the performance gap,this paper conducts research in two classic fields of crossdomain,namely unsupervised domain adaptation and incremental learning.We propose Connective Momentum Clustering(CMC)and Augmented Geometric Distillation(AGD)respectively.In CMC,we regard the annotations of the source domain as connective knowledge,that connections exist between data pairs which share the same identities.To transfer such connective knowledge to the target domain,we propose to leverage a proxy model,which is in term of Graph Convolutional Networks.It learns in source domain and inference reliably in targer domain with its unique capability to formulate graph pattern of neighborhoods in feature space.Additionally,to further improve transfering performance,we introduce momentum smoothing mechanism and normalization decoupling.Momentum smoothing acts as a low-pass filter to filter out noisy data,continually optimizing clustering quality.And normalization decoupling tackles BatchNorm pollution caused by mixed training of source and target domain data.On mainstream domain adaptation tasks Duke-to-Market and Market-toDuke,CMC achieves the performance of mAP=80.2%Rank-1=91.3%and mAP=70.4%Rank-1=82.4%,demonstrating advantage over other methods.In AGD,we first develop a general data-free distillation framework to handle the strict privacy in ReID.It generates dreaming memory through a pretrained prefix model to replay previous knowledge.However,the low quality of dreaming memory leads to the "noisy distillation".Towards this,we propose augmented distillation scheme,which distills knowledge in a novel criss-cross pattern,to highlight effective information and cancel out noise.In addition,to achieve a better trade-off between learning and memorizing,we propose geometric distillation loss.It departs from the traditional approach of penalizing semantic drift and introduces the novel idea of constraining geometric structure of feature space during its drift.On incremental tasks MSMT-to-Market and MSMT-to-PersonX,AGD yields performance of mAP=61.2%Rank-1=79.7%and mAP=61.4%Rank-1=79.9%,surpassing existing algorithms.Moreover,AGD exhibits impressive generalization in class incremental learning due to its plug-and-play nature.