Depth Perception in Frogs and Toads

In this study, we present a sequence of models exploring neural mechanisms by which frogs and toads may extract depth information from visual data. Although the models are constructed to be true, as far as possible, to what is known regarding depth perception in frogs and toads, they are also of more general interest. In particular, when they were developed they were the first depth models to exploit both binocular and monocular depth cues in a cooperative fashion.

In this chapter we review the reasons why previous theoretical models do not provide an adequate explanation for the depth perception process in frogs and toads. We begin with a survey of earlier depth models and a review of experimental evidence from behavioral, anatomical, and physiological experiments relating to depth perception in frogs and toads. We conclude by contrasting the assumptions made in the models with observations about the animals. We note that all of the previous depth models based on stereopsis are restricted to the consideration of binocular matching on a pair of static images, and that their purpose is to produce a depth-mapping from the image pair. These models also depend upon assumptions concerning vergence and image granularity. We show that these restrictions and assumptions are not applicable to the depth resolution problem in frogs and toads.

In this chapter we introduce a model of depth perception that explores one way in which lens accommodation cues can be used to help disambiguate the correspondence problem of stereopsis. Although the model was originally developed to explain how frogs and toads accomplish binocular depth perception without the use of cues from vergence, it is also of general interest as an extension to the class of cooperative stereo models. It consists of two cooperative depth discrimination processes, each acting to build a depth map, with one process receiving monocular depth cues based on accommodation and the other receiving binocular depth cues based on disparity. Mutual excitatory connections between the maps allow the model to converge to a single solution where accuracy is governed by binocularity. The model will also function in a purely monocular mode if binocular input is removed. Neurophysiological data on the visual system of frogs and toads are used to constrain choices made in constructing the model, and results obtained from simulation runs are compared with data from behavioral experiments.

In this chapter we present an action-oriented model of the spatial localization of prey by frogs and toads. Instead of building a global depth map, we propose that the goal of catching prey can lead a frog or toad to select a particular region of its visual world for special scrutiny. We suggest that the first step of the prey-catching sequence is not an overt movement, but a covert movement to adjust the accommodative state of the lenses and thus lock the visual apparatus on to a stimulus. We demonstrate how prey localization can be acheived rapidly and accurately by coupling prey-selection and lens-accommodation processes within a feedback loop. Information derived from prey selection supplies a setpoint for accommodation. In turn, adjustment of the lens modifies the visual input and can alter the prey selection process. The natural feedback of this goal-seeking system automatically corrects for the problem of ambiguity in binocular matching. We tie the model to the known anatomy, physiology, and behavior of frogs and toads, identifying brain regions that could provide the neural substrates necessary to support the model’s various functional stages. We also present experiments, with a computer simulation, that compare the model’s functioning with animal behavior.

In our introductory chapter we expressed the hope that, by modeling a natural system, we could make contributions to the fields of both brain theory and robotics. Specifically, we wished to contribute to the understanding of the depth perception process in frogs and toads and, at the same time, to suggest depth algorithms that would be useful in the design of robotic systems. But our initial hope must be tempered by the fact that the disciplines under consideration are both in early and exploratory stages. Except under highly constrained conditions, robots do not yet make successful use of sensory feedback to control their actions. Similarly, attempts to relate animal behavior to underlying neurophysiology and neuroanatomy entail the making of assumptions that are often untestable using current methods. It is clear that, at this stage of development, a study such as ours can make only preliminary claims.