Advanced

Data Augmentation by Using Object Shape Reconstruction

Granholm, Mats LU and Johannesson, Filip LU (2017) In Master's Theses in Mathematical Sciences FMA820 20171
Mathematics (Faculty of Engineering)
Abstract
Achieving high performance face recognition often requires large manually labeled training datasets. As such datasets can be difficult to obtain, we investigate whether smaller datasets can be augmented synthetically in order to increase performance.

We use 3D morphable models to create 3D reconstructions of faces from only a single image. The 3D reconstructions are used to render new face images in different poses in order to augment the original dataset. We also investigate whether generative adversarial networks (GANs) can be used to create completely synthetic training datasets for face recognition.

We show that recognition performance can be improved for non-frontal images when augmenting with similarly posed synthetic images.... (More)
Achieving high performance face recognition often requires large manually labeled training datasets. As such datasets can be difficult to obtain, we investigate whether smaller datasets can be augmented synthetically in order to increase performance.

We use 3D morphable models to create 3D reconstructions of faces from only a single image. The 3D reconstructions are used to render new face images in different poses in order to augment the original dataset. We also investigate whether generative adversarial networks (GANs) can be used to create completely synthetic training datasets for face recognition.

We show that recognition performance can be improved for non-frontal images when augmenting with similarly posed synthetic images. Quality over quantity is found to be one of the most important aspects of the synthesizing procedure, where few high quality synthetic images perform better than many low quality synthetic images. We conclude that if higher quality reconstructions are achieved, the performance could be further improved. For future work, GANs seem promising for the task at hand. (Less)
Please use this url to cite or link to this publication:
author
Granholm, Mats LU and Johannesson, Filip LU
supervisor
organization
course
FMA820 20171
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Morphable models, Generative adversarial networks, Data augmentation
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3326-2017
ISSN
1404-6342
other publication id
2017:E41
language
English
id
8914822
date added to LUP
2017-06-20 15:09:05
date last changed
2017-06-20 15:09:05
@misc{8914822,
  abstract     = {Achieving high performance face recognition often requires large manually labeled training datasets. As such datasets can be difficult to obtain, we investigate whether smaller datasets can be augmented synthetically in order to increase performance.

We use 3D morphable models to create 3D reconstructions of faces from only a single image. The 3D reconstructions are used to render new face images in different poses in order to augment the original dataset. We also investigate whether generative adversarial networks (GANs) can be used to create completely synthetic training datasets for face recognition.

We show that recognition performance can be improved for non-frontal images when augmenting with similarly posed synthetic images. Quality over quantity is found to be one of the most important aspects of the synthesizing procedure, where few high quality synthetic images perform better than many low quality synthetic images. We conclude that if higher quality reconstructions are achieved, the performance could be further improved. For future work, GANs seem promising for the task at hand.},
  author       = {Granholm, Mats and Johannesson, Filip},
  issn         = {1404-6342},
  keyword      = {Morphable models,Generative adversarial networks,Data augmentation},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences},
  title        = {Data Augmentation by Using Object Shape Reconstruction},
  year         = {2017},
}