The literature on computer vision contains hundreds of references to thinning or skeletonizing, but the only consistent definition for what is meant by a digital skeleton uses the medial axis transform. Most vision researchers would agree that the MAT does not yield an ideal, or in some cases even acceptable, skeleton. For example, single pixel irregularities can produce gross changes in an otherwise simple skeleton. The problem of defining what is meant by skeleton and skeletal pixel is one that has been rarely addressed, but seems crucial. Also, a great amount of past effort has gone into speeding up the thinning process without as much attention to the quality of the result, possibly because no good definition of a skeleton exists. By looking at the extensive literature on the subject some common properties of a good skeleton can be collected, but the process by which the skeleton is found should initially be ignored. This article will describe the properties of the skeleton of a binary object, both historically and by introducing new suggestions. Methods of producing skeletons from this definition will then be discussed, and simple metrics will be used to characterize the 'goodness' of the skeletons. The basic idea is that a skeleton is a global property of a binary object, and that the boundary should be used to locate the skeletal pixels.