Evaluating Test-Time Adaptation For Facial Expression Recognition Under Natural Cross-Dataset Distribution Shifts

A research team led by John Turnbull has published groundbreaking work on making facial expression recognition (FER) systems more robust in real-world conditions. Their paper, accepted at ICASSP 2026, presents the first comprehensive evaluation of Test-Time Adaptation (TTA) methods for FER under natural domain shifts. Unlike previous research focused on synthetic corruptions, this study examines real-world challenges caused by differing data collection protocols, annotation standards, and demographic variations across widely used FER datasets.

The research demonstrates that TTA methods can boost FER performance by up to 11.34% when models encounter natural distribution shifts. The team found that different TTA approaches excel in different scenarios: entropy minimization methods like TENT and SAR perform best with clean target distributions, prototype adjustment methods such as T3A handle larger distributional distances effectively, and feature alignment approaches like SHOT deliver the biggest gains when target distributions are noisier than source data. This cross-dataset analysis reveals that TTA effectiveness depends on both the distributional distance and the severity of natural shifts between domains.

This work represents a significant advancement in making AI systems more adaptable and reliable in practical applications. By enabling models to adapt during inference without requiring labeled source data or full retraining, TTA methods offer a practical solution to one of the most persistent challenges in computer vision deployment. The findings provide clear guidance for practitioners on which adaptation strategies to deploy based on the specific characteristics of their target environments.