Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Using transfer learning on 2D skeleton-based action recognition networks to improve performance on data from previously unseen camera angles

Cederquist, Filip LU and Edman Harrysson, Gustav LU (2021) In Master's Theses in Mathematical Sciences FMAM05 20202
Mathematics (Faculty of Engineering)
Abstract
Action recognition is a task in computer vision of inferring an action performed by a subject given an image or video. In this thesis we looked at how the performance of an action recognition network changed when introduced to new angles and how incorporating that data in the training affects this performance. A state-of-the-art skeleton extraction network HRNet\_w48+DARK was used together with a large action recognition dataset NTU RGB+D to create a baseline dataset containing skeleton data and actions labels. A new dataset was recorded which introduced a shift in vertical view point angle, which was used for analysis. Two different action recognition methods were used and compared to broaden the analysis. Results show that a small... (More)
Action recognition is a task in computer vision of inferring an action performed by a subject given an image or video. In this thesis we looked at how the performance of an action recognition network changed when introduced to new angles and how incorporating that data in the training affects this performance. A state-of-the-art skeleton extraction network HRNet\_w48+DARK was used together with a large action recognition dataset NTU RGB+D to create a baseline dataset containing skeleton data and actions labels. A new dataset was recorded which introduced a shift in vertical view point angle, which was used for analysis. Two different action recognition methods were used and compared to broaden the analysis. Results show that a small addition of 5-10$\%$ of the original amount of data from the new angles are enough to increase the accuracy on those angles greatly. The accuracy on the previous angles decreases but only by a small margin compared to the increased accuracy on the new angles. This implies that an extension of an action dataset to include another angle is a feasible task not requiring a large amount of data. (Less)
Please use this url to cite or link to this publication:
author
Cederquist, Filip LU and Edman Harrysson, Gustav LU
supervisor
organization
course
FMAM05 20202
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3441-2021
ISSN
1404-6342
other publication id
2021:E16
language
English
id
9061579
date added to LUP
2021-09-17 15:25:14
date last changed
2021-09-17 15:25:14
@misc{9061579,
  abstract     = {{Action recognition is a task in computer vision of inferring an action performed by a subject given an image or video. In this thesis we looked at how the performance of an action recognition network changed when introduced to new angles and how incorporating that data in the training affects this performance. A state-of-the-art skeleton extraction network HRNet\_w48+DARK was used together with a large action recognition dataset NTU RGB+D to create a baseline dataset containing skeleton data and actions labels. A new dataset was recorded which introduced a shift in vertical view point angle, which was used for analysis. Two different action recognition methods were used and compared to broaden the analysis. Results show that a small addition of 5-10$\%$ of the original amount of data from the new angles are enough to increase the accuracy on those angles greatly. The accuracy on the previous angles decreases but only by a small margin compared to the increased accuracy on the new angles. This implies that an extension of an action dataset to include another angle is a feasible task not requiring a large amount of data.}},
  author       = {{Cederquist, Filip and Edman Harrysson, Gustav}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master's Theses in Mathematical Sciences}},
  title        = {{Using transfer learning on 2D skeleton-based action recognition networks to improve performance on data from previously unseen camera angles}},
  year         = {{2021}},
}