A novel process used for selecting landscape scenes stored as multimedia objects within a video database is presented. First, each frame of a video clip undergoes a preprocessing step consisting of quantizing all actual image colors to a fixed set of 256 possible levels. Next, each frame of the video is segmented into realistic objects based on color/texture features. A distance measurement is then utilized for identifying those segmented objects that sufficiently match a preselected ground-truth objects. Each matched object is subsequently automatically annotated as being associated with a landscape scene. The original video, segmented objects, low-level texture/color features, along with the annotation are all stored as object-relational data within a video database system. Queries based on high level semantics are applied to the video database resulting in a more robust and more meaningful selection of video data. Results on a large range of video clips are provided.
展开▼
