Changing disparity (CD) and interocular velocity difference (IOVD) are two possible mechanisms for stereomotion perception. Researchers in neurophysiology and psychophysiology have been arguing for years over their existence and contributions to visual recognition. The previous models that relate both stereopsis and motion perception are not truly tuned to stereomotion velocity, nor do they consider the development of the model neurons. Thus, we seek to investigate the two mechanisms via computational models which can help us validate the hypotheses and make testable predictions. We are also interested in the information encoding and development process of the mechanisms in the neural system....[ Read more ]

Changing disparity (CD) and interocular velocity difference (IOVD) are two possible mechanisms for stereomotion perception. Researchers in neurophysiology and psychophysiology have been arguing for years over their existence and contributions to visual recognition. The previous models that relate both stereopsis and motion perception are not truly tuned to stereomotion velocity, nor do they consider the development of the model neurons. Thus, we seek to investigate the two mechanisms via computational models which can help us validate the hypotheses and make testable predictions. We are also interested in the information encoding and development process of the mechanisms in the neural system.

We have proposed two physiologically plausible models corresponding to CD and IOVD mechanisms. By applying the models directly to image sequences, we have shown that the single neuron is tuned to stereomotion velocity, and that the unknown input MID velocity can also be estimated from the population responses. We have compared the model responses with psychophysical experiments. From both the stereomotion direction discrimination based on a single CD neuron opponent energy and the stereomotion velocity estimation performance based on neural population responses, we have found that the model predictions suggest that the CD and IOVD mechanisms cannot be distinguished easily by breaking the correlations in binocularity or time in stimuli.

We have further extended the two models to two two-stage networks by replacing the designed filters in our constructed models with a set of adaptive basis vectors. By presenting the CD and IOVD networks with image sequences of moving natural stimuli and constraining the sparseness of the basis vectors, the networks can learn the stereomotion tuning with topological organization. Our results suggest that CD mechanism is more likely to be developed than IOVD mechanism.