Cast Search via Two-Stream Label Propagation

Jhih-Ciang Wu, Bing-Jhang Lin, Bing-Yuan Zeng, Li-Chen Fu, Chiou-Shann Fuh, Tyng-Luh Liu

Abstract: We address the problem of Cast Search by Portraits (CSP) where a facial portrait of a cast member is provided to retrieve from a given video clip those frames containing the query target. The underlying CSP formulation is related to the task of person re-identification. However, CSP is more challenging in that the provided query image is only a portrait of a certain cast member, and the instances of the target in the candidate video could have a very different visual appearance. Such drastic visual variations are not common in addressing the problem of person re-id. We propose a two-stream network architecture for tackling the CSP challenge and also participate in the public CSP competition. The overall outcome in the competition is promising and worth further effort to improve our proposed model.