In this paper we proposed a novel processing framework for elliptical object detection from single image. It is well known that single ellipse can be estimated from any combination of five boundary pixels. Our proposal is based on a Modified RANSAC, which performs detection by an iterative manner that randomly selecting five boundary pixels to construct an ellipse model; and evaluating the model's validity afterward. By taking into consideration that providing constraint for the sampling process can significantly improve the detection capability, toward the sampling process, we assign guidance by clustering boundary pixels according to a boundary curves' segment-reconnect strategy. In the experimental result section, we demonstrate that this proposal achieves advance in both computational efficiency and robustness, by comparing with RHT based method as well as human labeling results.