TOPSIS Aided Object Pose Tracking on RGB Images
In the problem of object pose estimation, one way to cope with the effect of ambiguity is to use multiple hypotheses. In this work, rather than generating the output pose based on a single object pose, our objective is to enable the system to be aware of the potential object ambiguity through mainta...
Main Authors: | , |
---|---|
Format: | Article |
Language: | English |
Published: |
IEEE
2023-01-01
|
Series: | IEEE Access |
Subjects: | |
Online Access: | https://ieeexplore.ieee.org/document/10348576/ |
_version_ | 1797376352096092160 |
---|---|
author | Mateusz Majcher Bogdan Kwolek |
author_facet | Mateusz Majcher Bogdan Kwolek |
author_sort | Mateusz Majcher |
collection | DOAJ |
description | In the problem of object pose estimation, one way to cope with the effect of ambiguity is to use multiple hypotheses. In this work, rather than generating the output pose based on a single object pose, our objective is to enable the system to be aware of the potential object ambiguity through maintaining multiple pose hypotheses. Firstly, we propose a pipeline for 6D object pose tracking on RGB images, wherein a key design is a fuzzy TOPSIS module that takes full advantage of multi-criteria decision making under uncertainties. Secondly, using decision variables determined on features that are frequently utilized in object pose estimation or tracking like segmented masks, fiducial keypoints, and distance transform the proposed method permits achieving tangible performance gains. An hourglass-based neural network is proposed to jointly detect object keypoints, predict the object’s non-occluded part, and to predict the object’s occluded part. To verify our designs, we conducted thorough experiments on the YCB-Video benchmark dataset. Besides, our method achieves competitive results in terms of ADD scores on the YCB-Video, showing that maintaining multiple pose hypotheses is beneficial to the task of object pose tracking. We observe that our method achieves competitive results against six recent methods estimating object pose from single frame and two SOTA object pose trackers. Extensive ablation studies verify our design choices. |
first_indexed | 2024-03-08T19:37:18Z |
format | Article |
id | doaj.art-65a76d85a916496997e33858d5a0da0f |
institution | Directory Open Access Journal |
issn | 2169-3536 |
language | English |
last_indexed | 2024-03-08T19:37:18Z |
publishDate | 2023-01-01 |
publisher | IEEE |
record_format | Article |
series | IEEE Access |
spelling | doaj.art-65a76d85a916496997e33858d5a0da0f2023-12-26T00:08:13ZengIEEEIEEE Access2169-35362023-01-011113949813950810.1109/ACCESS.2023.334091610348576TOPSIS Aided Object Pose Tracking on RGB ImagesMateusz Majcher0https://orcid.org/0000-0002-4444-7491Bogdan Kwolek1https://orcid.org/0000-0002-7715-1435Department of Computer Science, AGH University of Science and Technology, Kraków, PolandDepartment of Computer Science, AGH University of Science and Technology, Kraków, PolandIn the problem of object pose estimation, one way to cope with the effect of ambiguity is to use multiple hypotheses. In this work, rather than generating the output pose based on a single object pose, our objective is to enable the system to be aware of the potential object ambiguity through maintaining multiple pose hypotheses. Firstly, we propose a pipeline for 6D object pose tracking on RGB images, wherein a key design is a fuzzy TOPSIS module that takes full advantage of multi-criteria decision making under uncertainties. Secondly, using decision variables determined on features that are frequently utilized in object pose estimation or tracking like segmented masks, fiducial keypoints, and distance transform the proposed method permits achieving tangible performance gains. An hourglass-based neural network is proposed to jointly detect object keypoints, predict the object’s non-occluded part, and to predict the object’s occluded part. To verify our designs, we conducted thorough experiments on the YCB-Video benchmark dataset. Besides, our method achieves competitive results in terms of ADD scores on the YCB-Video, showing that maintaining multiple pose hypotheses is beneficial to the task of object pose tracking. We observe that our method achieves competitive results against six recent methods estimating object pose from single frame and two SOTA object pose trackers. Extensive ablation studies verify our design choices.https://ieeexplore.ieee.org/document/10348576/Object pose trackingmulti-criteria decision makingmulti-task neural networkshandling uncertain datafuzzy TOPSIS |
spellingShingle | Mateusz Majcher Bogdan Kwolek TOPSIS Aided Object Pose Tracking on RGB Images IEEE Access Object pose tracking multi-criteria decision making multi-task neural networks handling uncertain data fuzzy TOPSIS |
title | TOPSIS Aided Object Pose Tracking on RGB Images |
title_full | TOPSIS Aided Object Pose Tracking on RGB Images |
title_fullStr | TOPSIS Aided Object Pose Tracking on RGB Images |
title_full_unstemmed | TOPSIS Aided Object Pose Tracking on RGB Images |
title_short | TOPSIS Aided Object Pose Tracking on RGB Images |
title_sort | topsis aided object pose tracking on rgb images |
topic | Object pose tracking multi-criteria decision making multi-task neural networks handling uncertain data fuzzy TOPSIS |
url | https://ieeexplore.ieee.org/document/10348576/ |
work_keys_str_mv | AT mateuszmajcher topsisaidedobjectposetrackingonrgbimages AT bogdankwolek topsisaidedobjectposetrackingonrgbimages |