Friendly faces: Weakly supervised character identification

Abstract

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.