A method of long-tailed visual recognition via dynamic adjusting logit

Yaxuan Wang; Xinxi Lu; Yao Ning

doi:10.1117/12.3035311

19 July 2024 A method of long-tailed visual recognition via dynamic adjusting logit

Yaxuan Wang, Xinxi Lu, Yao Ning

Proceedings Volume 13213, International Conference on Image Processing and Artificial Intelligence (ICIPAl 2024); 132131E (2024) https://doi.org/10.1117/12.3035311
Event: International Conference on Image Processing and Artificial Intelligence (ICIPAl2024), 2024, Suzhou, China

Abstract

Extreme imbalance in long-tail datasets poses huge challenges for deep neural networks model and visual recognition. Networks models are very easy to be dominated by the "head" class and produce overfitting. At the same time, the ability of the model to model the "tail" class is extremely limited, which leads to serious model performance degradation. Most previous works attempt to address classes imbalance by assigning more weights to the tail classes, and these methods can cause the model to overfit the partial tail classes by directing the network to pay more attention to the tail classes. Although the model performance is improved, the embedding space of tail classes is not really expanded the input features are still biased, the problems of decision boundary shifting to tail classes and the tail class embedding space distortion are not solved. In order to address this problem, this paper proposes a Gaussian perturbation adjusted logit method to improve the decision boundary migration problem, and calibrates the embedding space distortion of tail class by giving different adjusted strength for different classes. At the same time, we study the adjusted strength and the correlation between the model performance, found that giving proper adjusted strength for different classification abilities could be a better solution for the class imbalance problems. The proposed method has been shown to be effective through our extensive experiments on popular long-tail datasets, and it has achievedstate of the art performance on multiple long-tail datasets.

(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Yaxuan Wang, Xinxi Lu, and Yao Ning "A method of long-tailed visual recognition via dynamic adjusting logit", Proc. SPIE 13213, International Conference on Image Processing and Artificial Intelligence (ICIPAl 2024), 132131E (19 July 2024); https://doi.org/10.1117/12.3035311

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available

Members: $17.00

Non-members: $21.00 ADD TO CART

PROCEEDINGS
8 PAGES

DOWNLOAD PAPER SAVE TO MY LIBRARY

GET CITATION

RIGHTS & PERMISSIONS

Get copyright permission Get copyright permission on Copyright Marketplace

KEYWORDS

Data modeling

Education and training

Head

Statistical modeling

Performance modeling

Visualization

Deep learning

Show All Keywords

Keywords/Phrases

Search In:

Publication Years