Survey on unsupervised learning methods for optical flow estimation

Tomislav Dobrički; Xiahai Zhuang; Kyoung Jae Won; Byung Woo Hong

doi:10.1109/ICTC55196.2022.9952910

Survey on unsupervised learning methods for optical flow estimation

Tomislav Dobrički^*, Xiahai Zhuang, Kyoung Jae Won, Byung Woo Hong

^*Corresponding author af dette arbejde

Won Group

Publikation: Bidrag til bog/antologi/rapport › Konferencebidrag i proceedings › Forskning › peer review

4 Citationer (Scopus)

Abstract

Optical flow is an important component in many computer vision applications. Thanks to deep learning, there have been great improvements in optical flow estimation in the past several years. But all of the top performing models are trained using a supervised method, on synthetic data sets. As the creation of accurately labeled optical flow data sets from real world images is incredibly difficult, many researchers have turned to developing unsupervised approaches. In this paper we conduct a survey of some of the most recent papers in unsupervised learning of optical flow, and present some of the key elements that are universally utilized. In addition, we did a results comparison, and found that the best performing unsupervised models are UnDAF-RAFT for the MPI-Sintel benchmark, and UpFlow on the KITTI benchmark. But both models still have considerably worse results when compared to supervised methods.

Originalsprog	Engelsk
Titel	ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence : Accelerating Digital Transformation with ICT Innovation
Antal sider	4
Forlag	IEEE Computer Society Press
Publikationsdato	2022
Sider	591-594
ISBN (Elektronisk)	9781665499392
DOI	https://doi.org/10.1109/ICTC55196.2022.9952910
Status	Udgivet - 2022
Begivenhed	13th International Conference on Information and Communication Technology Convergence, ICTC 2022 - Jeju Island, Sydkorea Varighed: 19 okt. 2022 → 21 okt. 2022

Konference

Konference	13th International Conference on Information and Communication Technology Convergence, ICTC 2022
Land/Område	Sydkorea
By	Jeju Island
Periode	19/10/2022 → 21/10/2022

Navn	International Conference on ICT Convergence
Vol/bind	2022-October
ISSN	2162-1233

Bibliografisk note

Funding Information:
This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the National Program for Excellence in SW(20170001000061001) supervised by the IITP (Institute of Information & communications Technology Planning & Evaluation) in 2022 and IITP-2021-0-01574, High-Potential Individuals Global Training Program and NRF-2020K2A9A2A06026659.

Publisher Copyright:
© 2022 IEEE.

Adgang til dokumentet

10.1109/ICTC55196.2022.9952910

Citationsformater

Dobrički, T., Zhuang, X., Won, K. J., & Hong, B. W. (2022). Survey on unsupervised learning methods for optical flow estimation. I ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence: Accelerating Digital Transformation with ICT Innovation (s. 591-594). IEEE Computer Society Press. https://doi.org/10.1109/ICTC55196.2022.9952910

Survey on unsupervised learning methods for optical flow estimation. / Dobrički, Tomislav; Zhuang, Xiahai; Won, Kyoung Jae et al.
ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence: Accelerating Digital Transformation with ICT Innovation. IEEE Computer Society Press, 2022. s. 591-594 (International Conference on ICT Convergence, Bind 2022-October).

Publikation: Bidrag til bog/antologi/rapport › Konferencebidrag i proceedings › Forskning › peer review

Dobrički, T, Zhuang, X, Won, KJ & Hong, BW 2022, Survey on unsupervised learning methods for optical flow estimation. i ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence: Accelerating Digital Transformation with ICT Innovation. IEEE Computer Society Press, International Conference on ICT Convergence, bind 2022-October, s. 591-594, 13th International Conference on Information and Communication Technology Convergence, ICTC 2022, Jeju Island, Sydkorea, 19/10/2022. https://doi.org/10.1109/ICTC55196.2022.9952910

Dobrički T, Zhuang X, Won KJ, Hong BW. Survey on unsupervised learning methods for optical flow estimation. I ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence: Accelerating Digital Transformation with ICT Innovation. IEEE Computer Society Press. 2022. s. 591-594. (International Conference on ICT Convergence, Bind 2022-October). doi: 10.1109/ICTC55196.2022.9952910

Dobrički, Tomislav ; Zhuang, Xiahai ; Won, Kyoung Jae et al. / Survey on unsupervised learning methods for optical flow estimation. ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence: Accelerating Digital Transformation with ICT Innovation. IEEE Computer Society Press, 2022. s. 591-594 (International Conference on ICT Convergence, Bind 2022-October).

@inproceedings{6706ecc101124cdba8aace727d937430,

title = "Survey on unsupervised learning methods for optical flow estimation",

abstract = "Optical flow is an important component in many computer vision applications. Thanks to deep learning, there have been great improvements in optical flow estimation in the past several years. But all of the top performing models are trained using a supervised method, on synthetic data sets. As the creation of accurately labeled optical flow data sets from real world images is incredibly difficult, many researchers have turned to developing unsupervised approaches. In this paper we conduct a survey of some of the most recent papers in unsupervised learning of optical flow, and present some of the key elements that are universally utilized. In addition, we did a results comparison, and found that the best performing unsupervised models are UnDAF-RAFT for the MPI-Sintel benchmark, and UpFlow on the KITTI benchmark. But both models still have considerably worse results when compared to supervised methods. ",

keywords = "deep learning, occlusion, Optical flow, photometric consistency, self-supervision, smooth regularization, unsupervised learning",

author = "Tomislav Dobri{\v c}ki and Xiahai Zhuang and Won, {Kyoung Jae} and Hong, {Byung Woo}",

note = "Publisher Copyright: {\textcopyright} 2022 IEEE.; 13th International Conference on Information and Communication Technology Convergence, ICTC 2022 ; Conference date: 19-10-2022 Through 21-10-2022",

year = "2022",

doi = "10.1109/ICTC55196.2022.9952910",

language = "English",

series = "International Conference on ICT Convergence",

publisher = "IEEE Computer Society Press",

pages = "591--594",

booktitle = "ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence",

address = "United States",

}

TY - GEN

T1 - Survey on unsupervised learning methods for optical flow estimation

AU - Dobrički, Tomislav

AU - Zhuang, Xiahai

AU - Won, Kyoung Jae

AU - Hong, Byung Woo

PY - 2022

Y1 - 2022

N2 - Optical flow is an important component in many computer vision applications. Thanks to deep learning, there have been great improvements in optical flow estimation in the past several years. But all of the top performing models are trained using a supervised method, on synthetic data sets. As the creation of accurately labeled optical flow data sets from real world images is incredibly difficult, many researchers have turned to developing unsupervised approaches. In this paper we conduct a survey of some of the most recent papers in unsupervised learning of optical flow, and present some of the key elements that are universally utilized. In addition, we did a results comparison, and found that the best performing unsupervised models are UnDAF-RAFT for the MPI-Sintel benchmark, and UpFlow on the KITTI benchmark. But both models still have considerably worse results when compared to supervised methods.

AB - Optical flow is an important component in many computer vision applications. Thanks to deep learning, there have been great improvements in optical flow estimation in the past several years. But all of the top performing models are trained using a supervised method, on synthetic data sets. As the creation of accurately labeled optical flow data sets from real world images is incredibly difficult, many researchers have turned to developing unsupervised approaches. In this paper we conduct a survey of some of the most recent papers in unsupervised learning of optical flow, and present some of the key elements that are universally utilized. In addition, we did a results comparison, and found that the best performing unsupervised models are UnDAF-RAFT for the MPI-Sintel benchmark, and UpFlow on the KITTI benchmark. But both models still have considerably worse results when compared to supervised methods.

KW - deep learning

KW - occlusion

KW - Optical flow

KW - photometric consistency

KW - self-supervision

KW - smooth regularization

KW - unsupervised learning

U2 - 10.1109/ICTC55196.2022.9952910

DO - 10.1109/ICTC55196.2022.9952910

M3 - Article in proceedings

AN - SCOPUS:85143256125

T3 - International Conference on ICT Convergence

SP - 591

EP - 594

BT - ICTC 2022 - 13th International Conference on Information and Communication Technology Convergence

PB - IEEE Computer Society Press

T2 - 13th International Conference on Information and Communication Technology Convergence, ICTC 2022

Y2 - 19 October 2022 through 21 October 2022

ER -