Quality-Aware Multimodal Fusion Reveals Implicit Identity in Valence-Arousal Features

Conventional face recognition relies on static appearance cues and degrades in unconstrained settings with expression variation, occlusion, and poor lighting. We hypothesize that audiovisual expression dynamics carry identity-discriminative information complementary to static appearance, and that extracting this signal requires multimodal representations robust to the variable input quality of in-the-wild video.

Summary generated by The Flow from the publisher's feed. The full article lives at Hugging Face Trending Papers.

• multimodal