StyleM: Stylized Metrics for Image Captioning Built with Contrastive N-grams

StyleCIDEr supports scoring the similarity of two compared captions with respect to their styles. We evaluate these two metrics using three stylized captioning methods trained on the PERSONALITY-CAPTIONS and FlickrStyle10K datasets: UPDOWN, MULTI-UPDOWN, and SVinVL. We also perform a human study to explore how well each caption aligns with human judgments in similar situations.