2019行人再识别年度进展回顾-valse.pptxVIP

行人再识别年度进展回顾;Person re-identification;Outline;Part-level features are effective.;PCB: Part-based Convolutional Baseline (Sun et al., ECCV2018);Dense semantic alignment (Zhang et al., CVPR2019);Similarity-guided GNN (Shen et al., ECCV18) absorbs information from neighboring pairs;Outline;; Image-image translation benefits UDA;Identity-preserving property benefits further.;Preserve two types of unsupervised similarities: self-similarity of an image before and after translation 2) domain-dissimilarity of a translated source image and a target image;Mining discriminative cues in the target domain improves the UDA accuracy;Learning a soft multi-label for unlabeled persons in target dataset by comparing to a set of known auxiliary reference persons;Video re-id: frame-level weights are important;Recurrent Convolutional Network (McLaughlin et al., CVPR 2016);Sequential Decision Making (Zhang et al., CVPR 2018 );Outline;End-to-end framework;Forward: each labeled identity is matched with all the stored features. Backward: Lookup Table is updated according to the id, pushing new features to Circular Queue, and pop out-of-date ones;Neural Person Search Machine (NPSM) (Liu et al., ICCV 2017);Separate person detector and re-identification seem to be a better choice.;PRW: DPM + IDE (Zheng et al., CVPR 2017) ;Open question: how does background affect the system?;Outline;Person re-id originates from person tracking;After years of study, person re-id features are improving tracking accuracies.;Re-id features are applied to multi-target multi-camera tracking;Siamese network, 49.3% MOTA （Ma et al., ACCV 2018）;Open question: will better re-id features lead to better tracking accuracy? ;Learning from synthetic data for person re-id;SOMAset and SOMAnet (Barbosa et al., CVIU 2018);SyRI (Bak et al., ECCV 2018);PersonX (Sun et al., CVPR 2019);DGNet (Zheng et al., CVPR 2019);Outline;Future research questions;Thank you!