中央研究院 資訊科學研究所

This talk concerns the classical problem of figure-ground segmentation from two different aspects. We first propose a meta-learning approach to figure-ground image segmentation. By exploring webly-abundant images of specific visual effects, our method can effectively learn the visual-effect internal representations in an unsupervised manner and uses this knowledge to capture the concept of figure and differentiate the figure from the ground in an image. In the second part of my talk, we cast the problem as referring image segmentation, in which a natural language referring expression is provided to guide pixel-level image segmentation. Referring image segmentation can be treated as richer-class semantic segmentation that poses technical challenges in CV and NLP. In both studies, we provide convincing SOTA experimental results to support the usefulness of our proposed methods.