BioKD Improves Video Emotion Recognition with Physiology.

Researchers have developed BioKD, a novel knowledge distillation framework designed to improve video-based emotion recognition, especially in situations where behavioral cues are ambiguous or socially masked. The framework addresses the challenge of integrating physiological signals, which are rich in emotional information but often noisy and difficult to deploy in real-world scenarios. BioKD uses these physiological signals as "privileged information" during the training phase to guide a video-based student model in learning deeper affective representations. To counter the inherent noise, inter-subject variability, and temporal inconsistencies of physiological data, BioKD incorporates a sample-wise reliability-aware gating mechanism alongside a progressive distillation strategy. This adaptive approach regulates the strength of knowledge transfer, effectively suppressing negative transfer from unreliable physiological supervision and ensuring more stable cross-modal distillation. Experiments on DEAP and AMIGOS datasets show BioKD consistently outperforms existing baselines for both valence and arousal recognition, particularly in challenging subject-independent evaluations, without introducing any additional inference-time overhead for the video student model.