Model reports

MER2026 MER-FG

原始来源英文简介

Track 2 of the MER2026 challenge: open-vocabulary, fine-grained emotion recognition from video, audio and subtitles, ranked on a 1,000-sample test set with emotion-wheel-based metrics that group synonyms (happy and joyful count as one label). Because the label space is open, how a system normalizes its vocabulary onto the wheel moves the score independently of recognition. Reference labels were built by merging top MER2025 team predictions and then manually validated.