Relevance feedback (RF) and object-oriented description for image content are efficient improvements for narrowing down the gap between low-level feature and high-level semantic concept in content-based image retrieval (CBIR). In this paper, a dynamic scheme for seeking user?s query concept is put forth in order to provide an accurate description for the query image content. To mine more semantic information accurately, an over-completed description space of image content based on region is first built. Furthermore, the non-parametric estimation is applied to modeling the over-completed description space with the premise of not knowing any prior knowledge about it. At the same time, the mean-shift optimization technique is adopted to grasp the semantic intent online in rounds of feedback. The experimental results on the Corel image database show that the proposed approach improves the retrieval performance largely.