A global parametric shape model (boundary) of the object is optimized according to evidence accumulated from local features and the prior probability of the model parameters learned from already segmented training samples. The parametric boundary is then deformed to the most conspicuous edge pixels near it. We emphasize the effectiveness of human intervention. Any unsatisfactory automatic segmentation is corrected interactively with mouse clicks. We demonstrate the above segmentation procedure on a flower image database of 1078 samples from 113 species. 257 out of 1078 images are segmented without interactive correction. On average, 5.7 mouse clicks are needed for each image, and the segmentation process takes 15.2 seconds.