Researchers have developed a novel audio-text system designed to recognize ambivalence and hesitancy in videos, specifically for the 11th ABAW Competition. This method processes videos in 5-second windows, integrating prosodic audio descriptors, RoBERTa embeddings, and handcrafted features that capture linguistic cues of uncertainty and conflict. The system employs temporal cross-attention for fusing audio and text, with support features influencing a gated multiple-instance learning pool. An ensemble of five models achieved a 0.875 average precision and a 0.72 macro-F1 score on the development set, and the source code is publicly available. AI
IMPACT Introduces a new approach for analyzing nuanced human emotions and communication styles in video content.
RANK_REASON Academic paper detailing a new method for a specific recognition task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
