VISAFF: Speaker-Centered Visual Affective Feature Learning for Emotion Recognition in Conversation

In three linesVISAFF is a framework for Emotion Recognition in Conversation (ERC) using vision-language models. It combines two stages: speaker-centered affective grounding and reliability-guided affective complementation. The tuning-free approach leverages frozen VLMs' reasoning capabilities, integrating visual, textual, and acoustic signals to improve accuracy without expensive fine-tuning.

Summary generated by Claude — human-verified