Data Ingestion, Analysis and Annotation, Second Version

D1.2: Data Ingestion, Analysis and Annotation, Second Version (CERTH: M20). Describes initial set of data ingestion services delivered to TVP, covering partner CMS/MAMS, social media, Web logs, video/TV viewing logs. Includes further developments and evaluation of improvements for video abstractions. Presents the updated and extended versions of concept-based video abstractions. Specifically, MTL and structured outputs will be introduced in our deep learning architectures targeting more accurate concept detection, improved fine-tuning algorithms will be developed that consider the commonalities between the source and target domain, the developed algorithms will be scaled to a bigger pool of target concepts. Includes further developments and evaluation of improvements for brand detection. Specifically, we will extend the brand detection algorithms with inductive transfer in our CNN architectures, more brand labels and advanced fusion methods for the final video-level brand detection.