EmoCue-Eval

One prerequisite assumption of EmoCue-360 is that the human-annotated data used for evaluation must be of sufficiently high quality and capable of covering all major emotional cues present in the video segments. At present, among publicly available datasets, only the EMER dataset can reasonably meet this requirement in terms of annotation quality and cue-level granularity. To further facilitate systematic evaluation research in the EMER task, we additionally construct a new benchmark, EmoCue-Eval, which contains 400 carefully and finely annotated test samples covering diverse scenarios, speakers, and emotional states. To the best of our knowledge, EmoCue-Eval is currently the largest EMER benchmark of its kind. This dataset provides a more systematic benchmark for future research on the comparison of EMER models.