An empirical-statistical agenda for recognition

This piece first describes what I see as the significant weaknesses in curr ent understanding of object recognition. We lack good schemes for: using un reliable information - like radiometric measurements - effectively; integra ting potentially contradictory cues; revising hypotheses in the presence of new information; determining potential representations from data; and supp ressing individual differences to obtain abstract classes. The problems are difficult, but none are unapproachable, given a change of emphasis in our research. All the important problems have a statistical flavour to them. Most involve a change of emphasis from the detailed study of specific cues to an invest igation of techniques for turning cues into integrated representations. In particular, all have a statistical flavour, and can be thought of as infere nce problems, I shaw an example that suggests that methods of Bayesian infe rence can be used to attack these difficulties. We have largely mapped out the the geometrical methods we need. Similarly, all the radiometric information that conceivably could be useful already ex ists. I believe that the next flowering of useful vision theories will occu r when we engage in an aggressive study of statistics and probabilistic mod elling, particularly methods of Bayesian inference.