Teaching with Soft Label Smoothing for Mitigating Noisy Labels in Facial Expressions

Lukov, Tohar; Zhao, Na; Lee, Gim Hee; Lim, Ser-Nam

doi:10.1007/978-3-031-19775-8_38

Cited by 8 publications

(3 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This conclusion is consistent with previous own conclusions in [63,149] and has also been discussed by others [15,38,177]. In addition, the use of label smoothing is an established technique to improve the training of neural networks [106,107,116]. The ambiguity in the data could be a possible reason why and when it works.…”

Section: Research Question 41: "Is One Annotation Enough To Capture T...supporting

confidence: 91%

“…For α = 1, the uniform probability distribution of 1/K is preferred over all previous distribution information of P(L x = k). This approach is successfully applied in many DL approaches to improve the overall performance [89,182], to mitigate annotation errors [107] or as a regularization [106]. The literature is divided on when or why label smoothing helps [116,192].…”

Section: Soft Labels and Label Smoothingmentioning

confidence: 99%

See 1 more Smart Citation

Addressing the Challenge of Ambiguous Data in Deep Learning

Schmarje

View full text Add to dashboard Cite

In machine learning, the availability of high-quality labeled data is essential for training accurate models. However, humans often disagree among themselves or over time when labeling or annotating image classification data. As a result, they create ambiguous data that poses a significant challenge to deep learning. This research focuses on addressing this issue by proposing an overview that defines the problem, provides the necessary data for research, establishes evaluation metrics and presents methods for solving the problem. The ultimate goal is to develop an effective annotation strategy for image classification based on several thousand different experiments to improve data quality while reducing cost. The main idea of this dissertation is not to rely on one annotation per image, but to consider the average of multiple annotations. However, this would not be feasible in reality without reducing the cost of acquiring such annotations. This dissertation uses proposals as a guide during the annotation process to improve the consistency of the annotations and to reduce the annotation time and thus the associated costs. The use of proposals could introduce a bias into the data. Analysis of this bias and its introduction led to possible methods to minimize or reverse it. All this research is finally unified in an annotation strategy in the form of a flowchart that allows future research to easily understand the necessary steps needed to produce high quality data for their specific use case with as little effort as possible. This strategy is successfully verified on a real biomedical task. In conclusion, this thesis presents a comprehensive investigation of the challenges posed by annotating ambiguous data in deep learning. This research provides an effective strategy to address the fundamental issue of data quality, thereby improving the quality and reliability of models and advancing the field of deep learning as a whole.

show abstract

Section: Research Question 41: "Is One Annotation Enough To Capture T...supporting

confidence: 91%

Section: Soft Labels and Label Smoothingmentioning

confidence: 99%

Addressing the Challenge of Ambiguous Data in Deep Learning

Schmarje

View full text Add to dashboard Cite

show abstract

“…Therefore, it is significant to build a robust Facial Expression Recognition (FER) System. In recent years, many FER methods [4,29,30,32,38,43,50,54] achieved state-of-the-art performance on several benchmark datasets (e.g. RAF-DB [31], SFEW [48] and AffectNet [34]).…”

Section: Introductionmentioning

confidence: 99%

Exploring Large-scale Unlabeled Faces to Enhance Facial Expression Recognition

Wang¹,

Cai²,

Li³

et al. 2023

Preprint

View full text Add to dashboard Cite

Facial Expression Recognition (FER) is an important task in computer vision and has wide applications in human-computer interaction, intelligent security, emotion analysis, and other fields. However, the limited size of FER datasets limits the generalization ability of expression recognition models, resulting in ineffective model performance. To address this problem, we propose a semisupervised learning framework that utilizes unlabeled face data to train expression recognition models effectively. Our method uses a dynamic threshold module (DTM) that can adaptively adjust the confidence threshold to fully utilize the face recognition (FR) data to generate pseudo-labels, thus improving the model's ability to model facial expressions. In the ABAW5 EXPR task, our method achieved excellent results on the official validation set.

show abstract