Advanced

Instance-level Object Detection of Grocery Store items using R-CNN:s

Rydén, Erik LU (2019) In Master's theses in mathematical sciences 2019:E15 FMA820 20171
Mathematics (Faculty of Engineering)
Abstract
Currently, the best ways to perform object detection in an image is to use a
neural network trained on large sets of annotated data. However, in certain
environments, the number of objects is large and gathering a sufficiently big
set of annotated data is prohibitively expensive. In this thesis, the possibility of using computer-generated artificial data instead in such an environment, is investigated. By training on computer generated images, the neural network may then be able to transfer this knowledge to the real world. More specifically, we investigate this method on the problem of detecting objects commonly found in grocery stores. This could then enable an efficient pipeline, in which every object would be photographed and... (More)
Currently, the best ways to perform object detection in an image is to use a
neural network trained on large sets of annotated data. However, in certain
environments, the number of objects is large and gathering a sufficiently big
set of annotated data is prohibitively expensive. In this thesis, the possibility of using computer-generated artificial data instead in such an environment, is investigated. By training on computer generated images, the neural network may then be able to transfer this knowledge to the real world. More specifically, we investigate this method on the problem of detecting objects commonly found in grocery stores. This could then enable an efficient pipeline, in which every object would be photographed and transferred into a textured 3D-model, which would then be used to generate the annotated images to train the network. This would remove the problem of gathering annotated
data. It is concluded that indeed, constitutes a possible approach to solving
the problem. (Less)
Please use this url to cite or link to this publication:
author
Rydén, Erik LU
supervisor
organization
course
FMA820 20171
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Image analysis, simulation, r-cnn
publication/series
Master's theses in mathematical sciences 2019:E15
report number
LUTFMA-3377-2019
ISSN
1404-6342
language
English
id
8975340
date added to LUP
2019-07-15 10:39:33
date last changed
2019-07-15 10:39:33
@misc{8975340,
  abstract     = {Currently, the best ways to perform object detection in an image is to use a
neural network trained on large sets of annotated data. However, in certain
environments, the number of objects is large and gathering a sufficiently big
set of annotated data is prohibitively expensive. In this thesis, the possibility of using computer-generated artificial data instead in such an environment, is investigated. By training on computer generated images, the neural network may then be able to transfer this knowledge to the real world. More specifically, we investigate this method on the problem of detecting objects commonly found in grocery stores. This could then enable an efficient pipeline, in which every object would be photographed and transferred into a textured 3D-model, which would then be used to generate the annotated images to train the network. This would remove the problem of gathering annotated
data. It is concluded that indeed, constitutes a possible approach to solving
the problem.},
  author       = {Rydén, Erik},
  issn         = {1404-6342},
  keyword      = {Image analysis,simulation,r-cnn},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's theses in mathematical sciences 2019:E15},
  title        = {Instance-level Object Detection of Grocery Store items using R-CNN:s},
  year         = {2019},
}