Abundance of A/V material requires automatic annotation – automatic annotators typically operate in isolation and do not consider the context of a media resource – insufficient annotation quality for many tasks – available technologies are complex to use and difficult to combine: expensive, especially for SME approach – integration and orchestration of extractors for various media types, capable of reusing their results – providing all necessary technologies for distributed cross-media extraction, metadata publishing, querying and recommendations – core software provided under business-friendly OSS licenses (building on OSS projects such as Apache Marmotta), allowing inclusion of OSS and closed- source extractors.
The main challenges for MICO are cross-media extraction: container tag extraction, language detection, scene detection, face detection, speech-music discrimination, ASR and NER – context-specific, dynamic orchestration and parametrization: config of the analysis graph and extractor parameters depend on the context – bridge the gap between high-level user queries and context-specific low-level queries – metadata model, query language and reco for media fragments, e.g. find similar topics based on key word cooccurrence in different video fragment transcriptions
News Media Showcase
InsideOut10, Italian SME with extensive experience on web publishing and media delivery platforms brings a number of challenging business cases to MICO such as – “Shoof”: citizen journalism; challenges: automatic QA of user-created videos, ASR, tampering detection, etc.- “Greenpeace News”: digital magazine; challanges: metadata enrichment for “cross-issue” reco and discovery paths, etc. includes e.g. queries: “point me to scenes within videos where a specific person is talking about a topic, and show me similar content on related topics” with
Snapshot Serengeti Showcase
Zooniverse, home to some of the largest and most successful citizen-science projects in the world brings the challenges of the Snapshot Serengeti project to MICO, such as animal detection and classification – visual analysis: blank image detection, pre-classification – tactual analysis of posts and discussions, and recommendation: identify user interest, emerging species, enrich species information, detect controversial topics, identity user communities (e.g. via vocabulary expertise), targeted education of users, harness user expertise