Person Re-Identification Yiheng Liu Outli line Background - PowerPoint PPT Presentation

Person Re-Identification Yiheng Liu

Outli line • Background • Image-Based Person Re-Identification • Partial Person Re-identification • Video-Based Person Re-Identification • Our Methods

Outli line • Background • Image-Based Person Re-Identification • Partial Person Re-identification • Video-Based Person Re-Identification • Our Methods

Person Re-id identif ification • Person re-identification aims to match persons across non-overlapping surveillance camera views. 数据 检测 跟踪 匹配 Images ： https://www.tugraz.at/institute/icg/research/team-bischof/lrs/downloads/prid11/

Datasets and Protocols ls Dataset ID Images Market1501 1501 32217 DukeMTMC-reID 1812 36441 Image Cumulated Matching Characteristics (CMC) ： Rank-1, Rank-5 • CUHK03 1467 13164 • mean Average Precision (mAP) MSMT17 4101 126441 iLIDS-VID 300 42495 Video PRID2011 934 24541 MARS 1261 1191003

Outli line • Background • Image-Based Person Re-Identification • Partial Person Re-identification • Video-Based Person Re-Identification • Our Methods

Learning Dis iscrim iminative Features wit ith Mult ltiple Granularitie ies for Person Re-Id Identification

Beyond Part Models ls: Person Retrie ieval l wit ith Refin ined Part Poolin ing (and A Strong Convolutional l Baseline)

Refin ined Part Poolin ing

Horiz izontal l Pyramid Matchin ing for Person Re-id identific ication

Learning In Incremental Trip iple let Margin for Person Re-id identific ication

Outli line • Background • Image-Based Person Re-Identification • Partial Person Re-identification • Video-Based Person Re-Identification • Our Methods

Perceive Where to Focus: Learning Vis isibilit ity-aware Part-le level Features for Partial l Person Re-id identif ification

Overall distance: Region feature:

• The training of region locator • Cross-entropy loss • Triplet loss

Outli line • Background • Image-Based Person Re-Identification • Partial Person Re-identification • Video-Based Person Re-Identification • Our Methods

STA: Spatial-Temporal l Attention for Large-Scale le Vid ideo-based Person Re-Id Identif ification Inter-Frame Regularization

Feature Fusion Strategy

THANK YOU FOR LISTENING