From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation

Link
Abstract

Deep-learning-based gaze estimation approaches often suffer from notable performance degradation in unseen target domains. One of the primary reasons is that the Fully Connected layer is highly prone to overfitting when mapping the high-dimensional image feature to 3D gaze. In this paper, we propose Analytical Gaze Generalization framework (AGG) to improve the generalization ability of gaze estimation models without touching target domain data. The AGG consists of two modules, the Geodesic Projection Module (GPM) and the Sphere-Oriented Training (SOT). GPM is a generalizable replacement of FC layer, which projects high-dimensional image features to 3D space analytically to extract the principle components of gaze. Then, we propose Sphere-Oriented Training (SOT) to incorporate the GPM into the training process and further improve cross-domain performances. Experimental results demonstrate that the AGG effectively alleviate the overfitting problem and consistently improves the cross-domain gaze estimation accuracy in 12 cross-domain settings, without requiring any target domain data. The insight from the Analytical Gaze Generalization framework has the potential to benefit other regression tasks with physical meanings.

Synth

Problem:: 고차원 Feature를 3D Gaze로 매핑하는 FC Layer의 Overfitting이 주요 원인 / FC Layer가 Domain-Specific한 비시선 정보까지 과도하게 학습함

Solution:: 고차원 Feature를 Geodesic Distance 기반 Isomap으로 3D 공간에 투영 / 투영된 3D Feature(PGF)를 적은 파라미터(10개)로 정렬 및 변환하여 Gaze 예측 / Feature Extractor를 PGF가 이상적 구면 분포에 가깝도록 추가 학습

Novelty:: Image Feature 간 Geodesic Distance와 실제 Gaze 각도 차이 간의 선형 비례 관계 발견 및 활용

Note:: Single-View Gaze Estimation이라는 좁은 분야에서 색다른 접근법을 도입한 흥미로운 연구 / ResNet50의 Base 성능(제안방식 X)이 너무 높아 확인할 필요가 있음 → Base 성능이 이미 SOTA임 / 24년도 CVPR인데 자기보다 성능이 좋은 SOTA는 다 비교에서 제외함 (대표적으로 CRGA)

Summary

Motivation

Method

Analytical Gaze Generalization (AGG) Framework

file-20250422165301024.png|725

Geodesic Projection Module (GPM)

Sphere-Oriented Training (SOT)

Method 검증