Causal Representation-Based Domain Generalization on Gaze Estimation

Link
Abstract

The availability of extensive datasets containing gaze information for each subject has significantly enhanced gaze estimation accuracy. However, the discrepancy between domains severely affects a model’s performance explicitly trained for a particular domain. In this paper, we propose the Causal Representation-Based Domain Generalization on Gaze Estimation (CauGE) framework designed based on the general principle of causal mechanisms, which is consistent with the domain difference. We employ an adversarial training manner and an additional penalizing term to extract domain-invariant features. After extracting features, we position the attention layer to make features sufficient for inferring the actual gaze. By leveraging these modules, CauGE ensures that the neural networks learn from representations that meet the causal mechanisms’ general principles. By this, CauGE generalizes across domains by extracting domain-invariant features, and spurious correlations cannot influence the model. Our method achieves state-of-the-art performance in the domain generalization on gaze estimation benchmark.

Synth

Problem:: 시선 추정 모델의 도메인 간 성능 저하 / 데이터셋 간 분포 차이 (appearance, illumination 등) / 눈 영역 의존성으로 인한 spurious correlation 취약성 / Domain Adaptation 방법들의 실용성 한계 (타겟 데이터 필요)

Solution:: Causal Mechanism 원칙 기반 domain-invariant representation 학습 / AugMix로 도메인 이동 시뮬레이션하여 causal/non-causal factors 분리 / Adversarial training으로 도메인 불변 특징 추출 / Factorization loss로 representation 독립성 보장 / Attention layer로 중요 causal factors 가중치 부여

Novelty:: 시선 추정 분야 최초 causality 개념 도입 / Causal mechanisms의 4가지 원칙 (CCP, Stability, Modularity, Causal Heterogeneity) 통합 적용 / Domain Generalization에서 SOTA 성능 (평균 6.92°) / 타겟 데이터 없이 cross-dataset generalization 달성

Note:: 1-Stage cross-dataset 실험에서 가장 단순하면서 효과적인 듯 / Causal Mechanism 적용 동기는 이해가 가는데, Factorization Loss만으로 달성 가능한지는 의문스러움 → 관련 논문을 봐야할 듯

Summary

Motivation

문제제기

기존 방법의 한계

Method

CauGE Framework 개요

file-20250602205609748.png|425

Causal Mechanism의 일반 원칙 적용

file-20250602204946778.png

1. Common Cause Principle (CCP)

2. Stability

3. Modularity

4. Causal Heterogeneity

구체적인 구현 방법

file-20250602205020883.png

1. Simulating Domain Shift

2. Adversarial Intervention Classification

3. Causal Mechanisms Modularization

4. Gaze Prediction with Attention Layer

Method 검증

실험 설정

정량적 성능 비교

Ablation Study

정성적 분석

Class Activation Map 시각화

file-20250602205420866.png

t-SNE Visualization

file-20250602205253575.png|650

추가 실험

Domain Shift Simulation 비교

Attention Layer 효과성