Abstract: Selecting the most representative view and dynamic viewpoint planning are desirable and challenging due to possible occlusions in the current view, especially for highly non-interactive situations. In this paper, we present a "goodness" measure of viewpoints selection designed for texture-based flow visualization and cluster optimal viewpoints for further exploration. Our optimal viewpoint measure considers not only flow behavior and structural information but also projected image quality. We extract a series of optimal viewpoints based on K-means clustering algorithm due to the slow frame-rates or non-interactive process. The experiment results demonstrate that this method can automatically and effectively achieve the optimal viewpoint.