News-MESI: A Dataset for Multimodal News Excerpt Segmentation and Identification

In complex long-term news videos, the fundamental component is the news excerpt which consists of many studio and interview screens. Spotting and identifying the correct news excerpt from such a complex long-term video is a challenging task. Apart from the inherent temporal semantics and the complex generic events interactions, the varied richness of semantics within the text and visual modalities further complicates matters. In this paper, we delve into the nuanced realm of video temporal understanding, examining it through a multimodal and multitask perspective. Our research involves presenting a more fine-grained challenge, which we refer to as M ultimodal News E xcerpt S egmentation and I dentification. The objective is to segment news videos into individual frame-level excerpts while accurately assigning elaborate tags to each segment by utilizing multimodal semantics. As there is an absence of multimodal fine-grained temporal segmentation dataset at present, we set up a new benchmark called News-MESI to support our research. News-MESI comprises over 150 high-quality news videos sourced from digital media, with approximately 150 hours in total and encompassing more than 2000 news excerpts. By annotating it with frame-level excerpt boundaries and an elaborate categorization hierarchy, this collection offers a valuable chance for multi-modal semantic understanding of these distinctive videos. We also present a novel algorithm employing coarse-to-fine multimodal fusion and hierarchical classification to address this problem. Extensive experiments are executed on our benchmark to show how the news content is temporally evolving in nature. Further analysis shows that multi-modal solutions are significantly superior to the single-modal solution.