The holy grail of multimedia indexing and retrieval is developing algorithms capable of imitating human abilities in distinguishing and recognising semantic concepts within the content, so that retrieval can be based on ”real world” concepts that come naturally to users. In this paper, we discuss an approach to using segmented video objects as the midlevel connection between low-level features and semantic concept description. In this paper, we consider a video object as a particular instance of a semantic concept and we model the semantic concept as an average representation of its instances. A system supporting object-based search through a test corpus is presented that allows matching presegmented objects based on automatically extracted lowlevel features. In the system, relevance feedback is employed to drive the learning of the semantic model during a regular search process.