Given a set of unsorted views captured in a wide area, an effective solution is proposed for image self-organization. The method starts with an initialization step where a small number of key frame pairs are selected to set up a global reference. Given a query image we automatically relate it to the existing key frames based on their pair-wise similarity evaluation. Four major enhancements are made in this step to achieve better performance. Firstly, a recently developed technique, SURF, is applied for robust feature detection. Secondly, an efficient coarse-to-fine matching strategy is implemented. Thirdly, an improved global representation is defined over each image for accurate and fast similarity evaluation. Finally, the method is constantly updated by adding more query images. Experiments were carried out to evaluate the performances of image self-organization by using a large number of images captured from our university's campus