In my thesis work, I present a paradigm for probabilistic object recognition. The paradigm is a variation of the alignment method due to Huttenlocher and Ullman, which in its most basic form is ``hypothesize a minimal number of matches between image and object features, compute the position of the object, project the object into the image, and verify.'' In my probabilistic paradigm, each stage uses statistical representations and probabilistic criteria. In addition, I consider the object modeling problem and provide a method for automatically acquiring these representations from real and synthetic images of an object. I believe that statistical distributions and prior probabilities provide a unifying framework for representing recognition constraints on object appearance. I feel that the statistical representation coupled with probabilistic optimization methods will lead to efficient and robust recognition in practice.