Injecting MLLM-generated multimodal reasoning traces into a fused audio-visual-text emotion model, plus a balanced dual-contrastive loss, raises average accuracy on MER2024 from 77.5 to 84.7 percent.
Balanced contrastive learning for long-tailed visual recognition
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Multimodal Video Emotion Recognition with Reliable Reasoning Priors
Injecting MLLM-generated multimodal reasoning traces into a fused audio-visual-text emotion model, plus a balanced dual-contrastive loss, raises average accuracy on MER2024 from 77.5 to 84.7 percent.