140 likes | 152 Views
This project focuses on the generation of tags and metadata, keyframe extraction for image search engines, video classification, clustering and summarization, and propagation of tags. It involves setting up a database from unstructured YouTube travel videos, annotation of shot boundaries, and the use of YouTube Downloader Tool for social media data collection. The project aims to integrate different expertise in image and video processing, audio signal processing, and text detection/recognition.
E N D
Goals Core partners: TUB (coord.), EPFL, TUD, QMUL Aims: • Generation of tags and metadata (signal processing and/or users’ annotation) • Keyframe extraction for image search engines • Video classification, clustering and summarization • Propagation of tags (object detection/ similarity)
Database Preparation: • Setup database from unstructured channel “Travel” on YouTube.com • 100 videos + affiliated metadata (keywords, comments, user information etc.) • Annotation of shot boundaries (by TUD and TUB)
Youtube Downloader Tool also used by the IRP “Social Media Data Collection“
Youtube Downloader Information also interesting for IRP “Social Media Data Collection“
Overview • Integration of different expertise: • Image search engines, audio signal processing, video signal processing and text detection/ recognition
Overview Video stream: Shot Subshot Subshot Key Frame • Temporal segmentation • Detection of hard cut, fade and dissolve • Subshot boundary detection • Special importance for single shot videos and for long shots • Detection of new visual content appearing through camera operations or undetected transitions. • Extraction of key frames • efficient key frame extraction as a handshake between existing image search engines and video domain
Future Work Key framing [TUB] + Object Replica Detection [EPFL] + Reranking [TUD] + others
Future Work • Audio-Visual Source Localization in Videos [EPFL] • localization can be used for automatic tag generation/propagation
Future Work • Quality Tags [EPFL] on Key Frames [TUB] • Key frames are used to produce quality tag by no reference video quality assessment.
Future Work 13 Multimodal Video Reranking [TU Delft] Task: Retrieval of shots treating a particular topic (i.e., a semantic theme) Already done: Search over speech transcripts is improved by pseudo-relevance feedback. Selection of feedback documents is informed by visual features. Future Work: Expand reranking ideas developed on VideoCLEF 2008 dataset to YouTube-like video collections. Reranked results list
Conclusion 14 • IRP continues until end May 2009 • So far very good research progress – new tools • Good integration of partners • Tools of much interest and use for other IRP‘s • Key framing will be finalized end of May • Clustering will continue in new IRP • Demos: www.nue.tu-berlin.de