What are Hard Samples for Gait Emotion Recognition?

Emotion recognition is a relevant capability for understanding human behavior, enabling interactive systems to support natural interaction with their users. Among the diverse ways the literature tackles this task, gait emotion recognition (extracting the perceived emotion from how someone walks) remains challenging due to limited data availability and inconsistent annotations that often overlap, leading to ambiguous scenarios and harming training and validation processes. Given this challenge, we pose the following question: What are hard samples for gait emotion recognition? To answer this question, we investigate sample complexity in gait-based emotion recognition and its impact on learning dynamics, focusing on ambiguous and hard-to-learn samples using dataset cartography techniques that map samples based on model correctness and confidence. Our findings suggest that easier samples are those that highlight clear emotional cues, as expected, whereas ambiguous and difficult samples include restrained movements, static frames, and annotation inconsistencies. These insights can inform dataset curation and model training strategies for more reliable gait emotion recognition.