Abstract

Facial age estimation remains challenging due to diverse aging patterns, facial appearances, and imaging conditions. Inspired by how humans compare within narrow or broad age ranges depending on prediction difficulty, we propose the Gaussian-attentive transformer (GATE), an uncertainty-aware framework that adaptively adjusts its comparison scope. GATE predicts the mean and variance of a Gaussian distribution to model data-dependent uncertainty and employs Gaussian-masked cross-attention to focus narrowly for confident cases and broadly for ambiguous ones. We develop three variants: GATE-s for single-domain learning, GATE-m for domain-aware multidomain learning, and GATE-u for domain-agnostic unified learning. Extensive experiments show that GATE achieves state-of-the-art accuracy across diverse benchmarks, with GATE-m improving both accuracy and efficiency through domain-adaptive representation and GATE-u ensuring robust generalization to unseen data. The source codes ar available at https://github.com/cwlee00/GATE.

Overview

Publication

Chaewon Lee, JunHyeok Heo and Chang-Su Kim, "GATE: Gaussian-Attentive Transformer for Uncertainty-Aware Age Estimation," accepted to Computer Vision and Pattern Recognition (CVPR) Findings , 2026.