2019年在合肥举办的Valse大会有关行人再识别PTT资料OutlineTrain/test on the same domainUnsupervised domain adaptationPerson searchOther problemsFuture research questionsPart-level features are effectivePCB: Part-based Convolutional Baseline sun et al, ECCv2018a column vectorNp branches ofFC layersaverage1X1 Convpooling凵■Stripe partitionHSimple and effective93. 8%R-1 on Market-1501convolutional layersinput image from backbone networktensor tcalumn vectors g column vectors hID predictionsHorizontal Pyramid FeaturesMultiple Granularities(Yang et al,, AAAl 2019)(Wang et al, ACM MM2019一5 oftmax256GMPPredictionsoftmax 2048间20480「=2「0=208Global Branch。…D[200k地5D中I Conv 1x1o…Part-2 Branch yFeature Maps9。…fu Jd.20482F4remk25620LP25621ResNet- 50 backbone(Before res_ conw_ 2)Ik25622Part-3 BranchPart ofConv 1x1res_ conwdx&conv 5x Pooling ReductionPose and 3d information are beneficialPose-Sensitive EmbeddingDense semantic alignmentSarfraz et al, CVPR2018)(Zhang et al, CVPR2019在Demo imagesUV space for 24 body partsWarping image pixels to UV spacePedestrian alignment networkUV space(Torso-part)DSAP-image(Zheng et al, TCSVT2018(Torso-part)Images U-map V-ma24 DSAP-imagesKalayeh et al CVPr 2018Yao et al. IEEE tiP 2019Neural architecture search?Smoothness in the feature space is beneficialGallery 1Gallery 1Graph edg二-2--二Gallery 2Gallery 2Similarity-guided GNN (Shen et al, ECCv18Probeabsorbs intormation trom neighboring pairsGallery 3Gallery 3Gallery 4allery 4Supervised Smoothed ManifoldSpectral Feature Transformation(Luo et al, 2018)Bai et al. CVPR2017considers the relations in the transformationenforces similarity smoothness among pairsdeep supervisionNX2048-dClassifiere0.85Spectral FeatureTransformationMetricembeddingAffinity GraphOff-lineN imagescosine→On-line256X128x3sImilaFeature Extracting Metric LearningAffinity Learning(Ours)Matching ProbabilityOutlineTrain/test on the same domainUnsupervised domain adaptationPerson searchOther problemsFuture research questionsEvolution in state-of-the-art PerformanceMarket-1501, Rank-1 accuracy8070605040。●Q自。●。。自鲁鲁。302002015(CVPR, ICCV2018(CVPR, ECCV)2019(CVPR)25%~35%35%~60%60%~75%Image-image translation benefits UDAIdentity-preserving property benefits further