Person Re-Identification in the Wild: From Short-Term to Long-Term

Huang, Yan

Person Re-Identification in the Wild: From Short-Term to Long-Term

Huang, Yan

Permalink

Publication Type:: Thesis
Issue Date:: 2021

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download contents and abstractAdobe PDF (343.78 kB)

Adobe PDF

Download thesisAdobe PDF (5.13 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Huang, Yan
dc.date.accessioned	2021-10-05T02:00:11Z
dc.date.available	2021-10-05T02:00:11Z
dc.date.issued	2021
dc.identifier.uri	http://hdl.handle.net/10453/150834
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	The task of person re-identification (re-ID) is to confirm the identity of a person in visual traces captured by different cameras. Person re-ID ``in the wild'' is a highly demanded technology and quite challenging due to the lack of data diversity, dramatic background variation between different domains, uncontrollable clothing change, and influence caused by the shortage of lighting. According to the scale of time gaps when footages are captured, these challenges are exposed into two different scenarios: 1) Short-Term person re-ID (ST-reID), and 2) Long-Term person re-ID (LT-reID). ST-reID addresses the time gap of several minutes. This scenario is mainly to deal with variations of illumination, viewpoint, and pose. Thus, data diversity is the primary concern. In addition, when footages are taken in different environments, ST-reID often encounters a large background shift issue. On the contrary, the time gaps between two footages in LT-reID can be several hours or even longer. Thus, a person has a great chance to change clothing. As another widely seen case, after a long-time gap, in order to take footages to re-identify a person when s/he reappears at night, infrared cameras are required. For ST-reID in the wild, the diversity of training data is essential to ensure a re-ID system can tolerate variations of illumination, viewpoint, and pose, etc. In addition, a model trained on one domain can lack certain generalization when it is applied to a new domain. This thesis will deeply study the two challenges exposed in ST-reID. Corresponding solutions are provided by using generated data to compensate for the limited data diversity. Also, a background shift suppression model is proposed to deal with the background shift issue for cross-domain ST-reID. For LT-reID in the wild, it is worth investigating approaches to tackle the clothing change issue. This thesis will introduce new clothing change datasets to the community. Corresponding solutions are given to tackle the clothing change issue. In addition, under tight security surveillance in LT-ReID, how to recognize the same person who appears under a RGB camera (in the daytime) and an infrared camera (reappear at night) is an immediate problem. A modality-biased training issue is unveiled for the infrared-visible LT-reID task, and corresponding solutions are given. This thesis will provide useful insights into diverse person re-ID issues in the wild from the short-term scenario to the long-term scenario to support practical usages in the real world.	en_US.UTF-8
dc.format	Thesis (PhD)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/150834/2/02whole.pdf
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	au.edu.uts.lib/ppc
dc.rights	info:eu-repo/semantics/openAccess
dc.title	Person Re-Identification in the Wild: From Short-Term to Long-Term	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

The task of person re-identification (re-ID) is to confirm the identity of a person in visual traces captured by different cameras. Person re-ID ``in the wild'' is a highly demanded technology and quite challenging due to the lack of data diversity, dramatic background variation between different domains, uncontrollable clothing change, and influence caused by the shortage of lighting. According to the scale of time gaps when footages are captured, these challenges are exposed into two different scenarios: 1) Short-Term person re-ID (ST-reID), and 2) Long-Term person re-ID (LT-reID). ST-reID addresses the time gap of several minutes. This scenario is mainly to deal with variations of illumination, viewpoint, and pose. Thus, data diversity is the primary concern. In addition, when footages are taken in different environments, ST-reID often encounters a large background shift issue. On the contrary, the time gaps between two footages in LT-reID can be several hours or even longer. Thus, a person has a great chance to change clothing. As another widely seen case, after a long-time gap, in order to take footages to re-identify a person when s/he reappears at night, infrared cameras are required. For ST-reID in the wild, the diversity of training data is essential to ensure a re-ID system can tolerate variations of illumination, viewpoint, and pose, etc. In addition, a model trained on one domain can lack certain generalization when it is applied to a new domain. This thesis will deeply study the two challenges exposed in ST-reID. Corresponding solutions are provided by using generated data to compensate for the limited data diversity. Also, a background shift suppression model is proposed to deal with the background shift issue for cross-domain ST-reID. For LT-reID in the wild, it is worth investigating approaches to tackle the clothing change issue. This thesis will introduce new clothing change datasets to the community. Corresponding solutions are given to tackle the clothing change issue. In addition, under tight security surveillance in LT-ReID, how to recognize the same person who appears under a RGB camera (in the daytime) and an infrared camera (reappear at night) is an immediate problem. A modality-biased training issue is unveiled for the infrared-visible LT-reID task, and corresponding solutions are given. This thesis will provide useful insights into diverse person re-ID issues in the wild from the short-term scenario to the long-term scenario to support practical usages in the real world.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/150834