The very first attempts at shape detection used entirely intensity based methods like thresholding. In thresholding , all intensities above a particular value are assumed to belong to one object and the rest belong to another. If there are multiple objects in the image, the intensity range is divided into that many bands. A lot of work has been done on smart methods for finding the threshold. These methods in general do not produce good results and have errors in localizing the boundary.
Model based techniques have also been used to perform shape detection. These involve the use of a template shape which is matched with every region in the image using correlation based approaches. The problem, of course, is that we need to have the shape in our model library.
Energy minimizing frameworks were proposed to solve the above problems. In these, an initial contour is specified, the contour than tries to move to a state which minimizes an energy term that depends on image features (intensity gradient, zero crossings, brightness) and contour features (continuity, curvature). A popular method that falls in this category is SNAKES which was introduced by Kass et al [2]. The technique uses marker points to describe the boundary and moves these points in a local neighborhood to a least energy position. These approaches , in general, can converge to any of the local minima and therefore require a good initial guess of the boundary to converge to the true boundary.
All of these techniques require the topology of the shapes to be known. They have problems detecting multiple objects, and have localization errors near sharp protrusions in the shape. The level set approach tries to overcome these problems.