Reuse of video annotations based on low-level descriptor similarity

The paper proposes a mixed annotation approach that exploits the advantages of both automatic and manual annotation techniques. Annotated multimedia material is regarded as a source of low- to high-level feature mappings supporting the propagation of annotations to new multimedia material. Video analysis tools do not currently produce effective annotations for retrieval, while manual annotation is expensive. The proposed approach uses low-level feature similarity to guide the retrieval of keyword annotations and aims to preserve the high quality of manual annotations while reducing the time and cost per annotated video unit. The annotation tool assists users, suggesting keywords for an item that come from similar items according to low-level descriptors. The effectiveness of current descriptors has been evaluated in an experimental environment using 5 video collections and a set of MPEG-7 descriptors. The similarity results have been compared to manually evaluated similarity.