Movie Character Retrieval with Few Examples

Movie character retrieval is a fundamental task in video understanding, with applications in video summarization and content-based retrieval. This paper presents an efficient system for retrieving character appearances in movies based on both facial and non-facial visual features. The proposed method extracts facial features using DeepFace with MTCNN for detection and FaceNet/ArcFace for embedding. For frames where faces are not detected, BEiT and CLIP models are used to extract contextual image features. The retrieval system is built using FAISS for fast and scalable indexing. Experimental results in the ACM Multimedia DVU Grand Challenges and TRECVID DVU 2023 datasets demonstrate the effectiveness of our approach, achieving a mean average precision (MAP) of 41.65% with optimal parameter tuning. The system exhibits robustness across different cinematic contexts and shows potential for scalability to large video datasets.