Cognitive Psychology Review MS # 99-068, Part 2
Computational Implications of Gestalt Theory II: A Directed Diffusion to Model Collinear Illusory Contour Formation
Author: Steven Lehar
This manuscript is alleged to be one of a two-part series on a computational Gestalt theory. I was given Part 2, only, to review. Part 2 makes frequent references to Part 1 which complicates the review of Part 1 in isolation, but probably not very much. This paper is a first for me in that the author cites the paper, itself, in the references, especially interesting in that the citation is for a different journal.
The paper provides a computational model for the second-order properties of illusory figures. Based on the citations, Part 1 was an account of the first-order properties of illusory contours. The particular computational model is contrasted to two other computation models of similar phenomena: one by Grossberg et al. and one by Zucker et al. The advantage claimed for the current model is that it can handle relatively complex phenomena without being prey to combinatorial explosion. The disadvantage admitted for the model is that it deliberately abandons neurological plausibility as a criterion for its components. Grossberg et al. worried a great deal about the neurological plausibility of their model. The defense for abandoning neural plausibility is that we don't know enough neuroscience yet and our understanding might well change.
The paper was well written and seemed fairly clear. I won't claim to be an expert on the computation details but I don't doubt the results of the simulations. I see the paper as providing an existence proof that a set of procedures that could produce something similar to illusory contours and could exhibit some of the properties of human perception. All that said, I don't see the paper as contributing to our understanding of illusory contours. My research group just spent a session on the recent Psch review paper by Roberts and Pashler on evaluating model fits. In my view, a model should be judged on it's ability to generate new ideas and a model isn't psychologically interesting unless it is testable. There was no hint in the paper of experiments that might be stimulated by the model or ways in which the model could be tested. Until those conditions are met, I would rather see such a paper published in a computer vision journal.