Advanced

Re-identification with Recurrent Neural Networks

Skog, Johannes LU (2017) In Master's Theses in Mathematical Sciences FMA820 20171
Mathematics (Faculty of Engineering)
Abstract
Today re-identification of persons in disjoint camera views demand large human effort and resources. There is a need of autonomous re-identification systems.

This thesis has been focused on the development of autonomous re-identification systems based on recurrent neural networks able to analyze sequences of images of persons, enabling the possibility to filter out outliers within sequences and make use of spatio-temporal features. Through the development and evaluation of multiple network architectures and recurrent units a broad examination of the viability of recurrent neural networks in re-identification systems have been made.

This thesis has shown, analysis of sequences of images can significantly increase the performance of... (More)
Today re-identification of persons in disjoint camera views demand large human effort and resources. There is a need of autonomous re-identification systems.

This thesis has been focused on the development of autonomous re-identification systems based on recurrent neural networks able to analyze sequences of images of persons, enabling the possibility to filter out outliers within sequences and make use of spatio-temporal features. Through the development and evaluation of multiple network architectures and recurrent units a broad examination of the viability of recurrent neural networks in re-identification systems have been made.

This thesis has shown, analysis of sequences of images can significantly increase the performance of re-identification systems compared to analysis of single images, with a doubling of the re-identification performance in multiple cases. Non-dynamic features tend to dominate the heap of information from sequences used during re-identification.

A realistic benchmark dataset was created as part of the thesis. Evaluation on the benchmark dataset has shown, to achieve satisfactory performance in real-world scenarios of autonomous re-identification systems, based on recurrent neural networks, larger training datasets with greater variety are needed. (Less)
Please use this url to cite or link to this publication:
author
Skog, Johannes LU
supervisor
organization
course
FMA820 20171
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3328-2017
ISSN
1404-6342
other publication id
2017:E54
language
English
id
8925200
date added to LUP
2017-09-19 16:02:11
date last changed
2017-09-19 16:02:11
@misc{8925200,
  abstract     = {Today re-identification of persons in disjoint camera views demand large human effort and resources. There is a need of autonomous re-identification systems.

This thesis has been focused on the development of autonomous re-identification systems based on recurrent neural networks able to analyze sequences of images of persons, enabling the possibility to filter out outliers within sequences and make use of spatio-temporal features. Through the development and evaluation of multiple network architectures and recurrent units a broad examination of the viability of recurrent neural networks in re-identification systems have been made.

This thesis has shown, analysis of sequences of images can significantly increase the performance of re-identification systems compared to analysis of single images, with a doubling of the re-identification performance in multiple cases. Non-dynamic features tend to dominate the heap of information from sequences used during re-identification.

A realistic benchmark dataset was created as part of the thesis. Evaluation on the benchmark dataset has shown, to achieve satisfactory performance in real-world scenarios of autonomous re-identification systems, based on recurrent neural networks, larger training datasets with greater variety are needed.},
  author       = {Skog, Johannes},
  issn         = {1404-6342},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences},
  title        = {Re-identification with Recurrent Neural Networks},
  year         = {2017},
}