Developing vision

Human infants are born with limited cognitive and perceptual abilities, yet within a few years their abilities outstrip those of any artificial intelligence program. We propose that studying the developmental process of the human visual system provides useful insights for creating better learned computer vision systems. We present four experiments to demonstrate this theory.; Our first experiment considers the hypothesis that systems learning binocular disparity estimation may benefit from the use of developmental progressions. We compare the performance of four models. Three were “developmental models” they received a relatively impoverished input early in training and the quality of which improved during the course of training. One model used a coarse-scale-to-multiscale developmental progression, another used a fine-scale-to-multiscale progression, and the third used a random progression. The final model was non-developmental: the nature of its input remained the same throughout training. The simulation results show that the coarse-scale-to-multiscale and the fine-scale-to-multiscale models performed best.; Our second experiment applied the same hypothesis to the problem of motion velocity estimation. The same four models were trained to estimate the velocity of an object. The simulation results show that the coarse-scale-to-multiscale model performed best.; Our third experiment examines the notion of modularity in multiple cue perceptual problems. In this experiment, we use motion and stereo information to estimate visual depth. We compare the performance of two models: a modular model where the two cues are separated and have limited interaction; and a monolithic model where the two cues have unlimited interaction. Simulation results show that the modular model performed best.; Our final experiment tests the hypothesis that more reliable cues can be used to teach less reliable cues in multiple cue perception problems. In this experiment we compare the performance of a sequentially trained model, which was first trained with motion data, and then trained with stereo data, to a model only trained with stereo data on estimation of visual depth from stereo data. Results show that the sequential model sometimes shows significant improvement over the stereo model.; We conclude that learned computer vision systems can benefit from the study of human perceptual development.