A Modified Self-training Method for Adapting Domains in the Task of Food Classification

Elnaz Jahani Heravi, Hamed Habibi Aghdam, Domenec Puig

Research output: Chapter in Book/Report/Conference proceedingConference contributionResearchpeer-review

Abstract

Food trackers are tools that recognize foods using their images. In the core of these tools there is usually a neural network that performs the classification. Neural networks are highly expressive models that need a large dataset to generalize well. Since it is hard to collect a training set that captures most of realistic situations in real world, there is usually a shift between the training set and the actual test set. This potentially reduces the performance of the network. In this paper, we propose a method based on self-training to perform unsupervised domain adaptation in the task of food classification. Our method takes into account the uncertainty of predictions instead of probability scores to assign pseudo-labels. Our experiments on the Food-101 and the UPMC-101 datasets show that the proposed method produces more accurate results compared to Tri-training method which had previously surpassed other domain adaptation methods.
Original languageEnglish
Title of host publicationVISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision
EditorsAlain Tremeau, Giovanni Maria Farinella, Jose Braz
PublisherSciTePress
Pages143-154
Volume5
ISBN (Electronic) 978-989758354-4
DOIs
Publication statusPublished - 2019
Externally publishedYes
EventVISIGRAPP 2019: 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Prague, Czech Republic
Duration: 25 Feb 201927 Feb 2019

Conference

ConferenceVISIGRAPP 2019
CountryCzech Republic
CityPrague
Period25/02/1927/02/19

Fingerprint

Neural networks
Labels
Experiments
Uncertainty

Cite this

Jahani Heravi, E., Habibi Aghdam, H., & Puig, D. (2019). A Modified Self-training Method for Adapting Domains in the Task of Food Classification. In A. Tremeau, G. M. Farinella, & J. Braz (Eds.), VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision (Vol. 5, pp. 143-154). SciTePress. https://doi.org/10.5220/0007688801430154

A Modified Self-training Method for Adapting Domains in the Task of Food Classification. / Jahani Heravi, Elnaz; Habibi Aghdam, Hamed; Puig, Domenec.

VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision. ed. / Alain Tremeau; Giovanni Maria Farinella; Jose Braz. Vol. 5 SciTePress, 2019. p. 143-154.

Research output: Chapter in Book/Report/Conference proceedingConference contributionResearchpeer-review

Jahani Heravi, E, Habibi Aghdam, H & Puig, D 2019, A Modified Self-training Method for Adapting Domains in the Task of Food Classification. in A Tremeau, GM Farinella & J Braz (eds), VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision. vol. 5, SciTePress, pp. 143-154, VISIGRAPP 2019, Prague, Czech Republic, 25/02/19. https://doi.org/10.5220/0007688801430154
Jahani Heravi E, Habibi Aghdam H, Puig D. A Modified Self-training Method for Adapting Domains in the Task of Food Classification. In Tremeau A, Farinella GM, Braz J, editors, VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision. Vol. 5. SciTePress. 2019. p. 143-154 https://doi.org/10.5220/0007688801430154
Jahani Heravi, Elnaz ; Habibi Aghdam, Hamed ; Puig, Domenec. / A Modified Self-training Method for Adapting Domains in the Task of Food Classification. VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision. editor / Alain Tremeau ; Giovanni Maria Farinella ; Jose Braz. Vol. 5 SciTePress, 2019. pp. 143-154
@inproceedings{317f9b3ae3db4f7c82eadfb715d55f6c,
title = "A Modified Self-training Method for Adapting Domains in the Task of Food Classification",
abstract = "Food trackers are tools that recognize foods using their images. In the core of these tools there is usually a neural network that performs the classification. Neural networks are highly expressive models that need a large dataset to generalize well. Since it is hard to collect a training set that captures most of realistic situations in real world, there is usually a shift between the training set and the actual test set. This potentially reduces the performance of the network. In this paper, we propose a method based on self-training to perform unsupervised domain adaptation in the task of food classification. Our method takes into account the uncertainty of predictions instead of probability scores to assign pseudo-labels. Our experiments on the Food-101 and the UPMC-101 datasets show that the proposed method produces more accurate results compared to Tri-training method which had previously surpassed other domain adaptation methods.",
author = "{Jahani Heravi}, Elnaz and {Habibi Aghdam}, Hamed and Domenec Puig",
year = "2019",
doi = "10.5220/0007688801430154",
language = "English",
volume = "5",
pages = "143--154",
editor = "Alain Tremeau and Farinella, {Giovanni Maria} and Jose Braz",
booktitle = "VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision",
publisher = "SciTePress",
address = "Portugal",

}

TY - GEN

T1 - A Modified Self-training Method for Adapting Domains in the Task of Food Classification

AU - Jahani Heravi, Elnaz

AU - Habibi Aghdam, Hamed

AU - Puig, Domenec

PY - 2019

Y1 - 2019

N2 - Food trackers are tools that recognize foods using their images. In the core of these tools there is usually a neural network that performs the classification. Neural networks are highly expressive models that need a large dataset to generalize well. Since it is hard to collect a training set that captures most of realistic situations in real world, there is usually a shift between the training set and the actual test set. This potentially reduces the performance of the network. In this paper, we propose a method based on self-training to perform unsupervised domain adaptation in the task of food classification. Our method takes into account the uncertainty of predictions instead of probability scores to assign pseudo-labels. Our experiments on the Food-101 and the UPMC-101 datasets show that the proposed method produces more accurate results compared to Tri-training method which had previously surpassed other domain adaptation methods.

AB - Food trackers are tools that recognize foods using their images. In the core of these tools there is usually a neural network that performs the classification. Neural networks are highly expressive models that need a large dataset to generalize well. Since it is hard to collect a training set that captures most of realistic situations in real world, there is usually a shift between the training set and the actual test set. This potentially reduces the performance of the network. In this paper, we propose a method based on self-training to perform unsupervised domain adaptation in the task of food classification. Our method takes into account the uncertainty of predictions instead of probability scores to assign pseudo-labels. Our experiments on the Food-101 and the UPMC-101 datasets show that the proposed method produces more accurate results compared to Tri-training method which had previously surpassed other domain adaptation methods.

U2 - 10.5220/0007688801430154

DO - 10.5220/0007688801430154

M3 - Conference contribution

VL - 5

SP - 143

EP - 154

BT - VISIGRAPP 2019 -Proceedings of the 14th International Joint Conference on Computer Vision

A2 - Tremeau, Alain

A2 - Farinella, Giovanni Maria

A2 - Braz, Jose

PB - SciTePress

ER -