Related Experiment Video
Updated: Jun 23, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
9.0K
Multitask Learning Strategy with Pseudo-Labeling: Face Recognition, Facial Landmark Detection, and Head Pose
Yongju Lee1, Sungjun Jang1, Han Byeol Bae2
1School of Electrical and Electronic Engineering, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul 03722, Republic of Korea.
Sensors (Basel, Switzerland)
|May 25, 2024
Summary
This study introduces a novel pseudo-labeling technique and multitask learning framework to improve facial analysis in real-world conditions. The method enhances facial landmark detection, head pose estimation, and face recognition, achieving state-of-the-art performance.
Area of Science:
- Computer Vision and Machine Learning
- Artificial Intelligence for Biometrics
Background:
- Facial analysis models often fail in real-world scenarios due to limitations in learning diverse human features and background noise.
- Existing training datasets for facial landmark detection and head pose estimation are frequently limited and noisy, hindering generalization.
- A significant gap exists between the performance of facial analysis models in standardized tests versus real-world applications.
Purpose of the Study:
- To bridge the performance gap between standardized and real-world facial analysis testing.
- To enhance the robustness and accuracy of facial landmark detection, head pose estimation, and face recognition.
- To develop a framework that overcomes limitations of diverse training data in facial analysis tasks.
Main Methods:
- Proposed a pseudo-labeling technique utilizing a diverse face recognition dataset to augment training data.
- Developed an integrated framework employing complementary multitask learning for robust feature extraction.
- Combined pseudo-labeling with multitask learning to promote the learning of pose-invariant features for improved face recognition.
Main Results:
- Achieved state-of-the-art (SOTA) or near-SOTA performance on AFLW2000-3D and BIWI datasets for facial landmark detection and head pose estimation.
- Demonstrated competitive face verification performance on the IJB-C dataset.
- Showcased stable performance even with training datasets lacking diverse face identifications, validated through a novel soft, medium, and hard case categorization.
Conclusions:
- The proposed pseudo-labeling and multitask learning framework significantly improves facial analysis performance in challenging, real-world conditions.
- The method effectively addresses the lack of data diversity and noise issues in training datasets.
- The integrated approach leads to more robust and generalizable facial analysis systems.
