Abstract: Techniques are disclosed for generating landmark models based on exemplar portions of images (“patches”) that are known to include the target landmark (forming a “positive” set of landmark vectors) and those that do not (forming a “negative” set of landmark vectors). Theses sets of positive and negative landmark vectors, along with other landmark statistics, form a landmark model. When an unknown image is received, candidate landmark vectors may be generated based on the image's content (or portion thereof) and applied to the landmark model to rapidly reduce the dimensionality (complexity) of the candidate landmark vector space. Landmark models may be applied to the reduced-dimensioned candidate landmark vectors to identify the most likely point corresponding to the target landmark.