Cue interaction in object recognition

We describe an experimental framework for a quantitative study of object recognition. Devising stimuli for 3-D object recognition that are both easily controlled in their complexity and sufficiently realistic is rather difficult. In this study we use a computer graphics system (Symbolics 3600) that permits us to overcome this difficulty. Stimulus complexity is controlled by using wirelike objects created by a random walk in 3-D and governed by a single parameter which reflects the complexity of the resulting structure. Our system allows us to investigate different theoretical approaches to object representation for recognition. We compare representations based on 3-D models, sets of 2½D canonical views, and a small number of distinctive 2-D features. Partial results of our study indicate that subjects tend to rely in their response, whenever possible, on local, 2-D features of the objects. Moreover, the choice of such features varies among subjects, as it appears from both reaction time and subjective judgment experiments.