Cross-modal correlation mining using graph algorithms

Multimedia objects like video clips or captioned images contain data of various modalities such as image, audio, and transcript text. Correlations across different modalities provide information about the multimedia content, and are useful in applications ranging from summarization to semantic captioning. We propose a graph-based method, MAGIC, which represents multimedia data as a graph and can find cross-modal correlations using "random walks with restarts." MAGIC has several desirable properties: (a) it is general and domain-independent; (b) it can detect correlations across any two modalities; (c) it is insensitive to parameter settings; (d) it scales up well for large datasets; (e) it enables novel multimedia applications (e.g., group captioning); and (f) it creates opportunity for applying graph algorithms to multimedia problems. When applied to automatic image captioning, MAGIC finds correlations between text and image and achieves a relative improvement of 58% in captioning accuracy as compared to recent machine learning techniques.