Earticle

초록

영어

Gaze estimation is essential in human–computer interaction, autonomous driving, and medical diagnosis, yet it suffers from susceptibility to environmental variations and poor generalization to unseen users. To address these challenges, we employ mixed-effects neural networks that capture both population-level (fixed) and individual-specific (random) patterns. However, existing models often rely on Gaussian assumptions, rendering them sensitive to outliers and incapable of inferring random effects for unseen clusters. We propose a t-distribution–based robust adversarially-regularized mixed effects neural network (RARMENN). The t-distribution enhances robustness against extreme values, while an adversarial classifier removes cluster-specific information from fixed effects, thereby enabling random-effect inference for new users. Experimental results on the MPIIGaze and UTMultiview datasets demonstrate up to a 16% improvement in within-subject accuracy and a 75% enhancement in LOOCV compared to Gaussian baselines, confirming significantly improved generalization and reliability for real-world gaze estimation.

목차

Abstract
1. 머리말
2. 혼합효과 신경망 모델
2.1. 선형 혼합효과 모델
2.2. 비선형 혼합효과 모델
3. ARMED 모델
3.1. 고정효과 하위네트워크
3.2. 변량효과 하위네트워크
3.3. 고정효과와 변량효과의 결합
3.4. 미학습 군집 예측
4. RARMENN 모델
4.1. t-분포 기반 고정효과 하위네트워크
4.2. t-분포 기반 변량효과 하위네트워크
4.3. 적대적 분류기 및 미학습 군집 대응
5. 실험 및 결과
5.1. 데이터셋 및 실험 설정
5.2. 실험 결과
6. 결론
Acknowledgments
참고문헌

키워드

저자

• 문동규 [ Dongkyu Moon | 대교 CNS 정보기술연구소 ]
• 심주용 [ Jooyong Shim | 인제대학교 통계학과 ]
• 황창하 [ Changha Hwang | 단국대학교 통계데이터사이언스학과 ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  통합자연과학논문집(구 조선자연과학논문집) [Journal of Integrative Natural Science]

• 간기

  계간

• pISSN

  2005-1042

• 수록기간

  2008~2026

• 등재여부

  KCI 등재

• 십진분류

  KDC 405 DDC 505