Motivation
nProblems in news video retrieval
nPrimary source of semantics
come from ASR, but ASR tends to be erroneous &
non-grammatical
nText does not necessarily
relate well to visual information
nLow-level features are
unreliable and unpredictable
nWhat we have:
nAnnotation of relevant high-level features (HLFs) with varying accuracy
nQuestion: How to
capitalize these HLF’s to support retrieval?