This paper demonstrates a novel method for automatically discovering and recognising characters in video without any labelled examples or user intervention. Instead weak supervision is obtained via a rough script-to-subtitle alignment. The technique uses pose invariant features, extracted from detected faces and clustered to form groups of co-occurring characters. Results show that with 9 characters, 29% of the closest exemplars are correctly identified, increasing to 50% as additional exemplars are considered.
CITATION STYLE
Marter, M., Hadfield, S., & Bowden, R. (2015). Friendly faces: Weakly supervised character identification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 8912, pp. 121–132). Springer Verlag. https://doi.org/10.1007/978-3-319-13737-7_11
Mendeley helps you to discover research relevant for your work.