Researchers have introduced Video2Reaction, a new multimodal dataset and benchmark designed to predict audience emotional responses to video content. The dataset, comprising over 10,000 videos, utilizes a two-stage pipeline with open-source LLMs to annotate induced emotions from social media, achieving 86% correctness. While finetuned foundation video models show promise, even state-of-the-art methods like LLaVA-NeXT struggle with the inherent subjectivity, achieving only 77% Top-3 F1 in dominant reaction prediction. AI
IMPACT This dataset could advance research in video understanding and personalized content recommendation by enabling better prediction of audience emotional responses.
RANK_REASON The cluster describes a new academic paper introducing a dataset and benchmark for a specific AI task.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
