Reproduce “Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wen Yin, Yong Wang, Guiduo Duan, Dongyang Zhang, Xin Hu, Yuan-Fang Li, Tao He

This paper introduces the Knowledge-aligned Counterfactual-enhancement Diffusion Perception (KCDP) framework for Unsupervised Cross-Domain Visual Emotion Recognition (UCDVER), a novel task for recognizing emotions across visually diverse domains without target-domain labels. KCDP achieved up to 14% higher accuracy compared to prior state-of-the-art domain adaptation models and an 11.39% lead over previous models on the Emo8 dataset.

Key insights

  • Structured, domain-agnostic knowledge triplets derived from image captions can effectively guide latent diffusion models to enhance visual features, making them more relevant to affective cues across domains.
  • Counterfactual reasoning, through interventions on emotional prompts and causal graph analysis, can generate robust pseudo-labels for unsupervised adaptation in abstract tasks like emotion recognition by maximizing emotion-specific alignment.
  • A Mixture of Experts (MoE) architecture successfully integrates diverse multimodal features (visual and knowledge representations), allowing for more robust cross-domain emotion classification than a single global classifier.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.