Keywords
neuromorphic computer vision, neural architecture, population coding, binocular energy model, disparity estimation
Abstract
The primate visual system efficiently and effectively solves a multitude of tasks from orientation detection to motion detection. The Computer Vision community is therefore beginning to implement algorithms that mimic the processing hierarchies present in the primate visual system in the hope of achieving flexible and robust artificial vision systems. Here, we reappropriate the neuroscience “borrowed” by the Computer Vision community and ask whether neuromorphic computer vision solutions may give us insight into the functioning of the primate visual system. Specifically, we implement a neuromorphic algorithm for disparity estimation and compare its performance against that of human observers. The algorithm greatly outperforms human subjects when tuned with parameters to compete with non-neural approaches to disparity estimation on benchmarking stereo image datasets. Conversely, when the algorithm is implemented with biologically plausible receptive field sizes, spatial selectivity, phase tuning, and neural noise, its performance is directly relatable to that of human observers. The receptive field size and the number of spatial scales sensibly determine the range of spatial frequencies in which the algorithm successfully operates. The algorithm’s phase tuning and neural noise in turn determine the algorithm’s peak disparity sensitivity. When included, retino-cortical mapping strongly degrades disparity estimation in the model’s periphery, further closening human and algorithm performance. Hence, a neuromorphic computer vision algorithm can be reappropriated to model human behavior, and can provide interesting insights into which aspects of human visual perception have been or are yet to be explained by vision science.
Start Date
13-5-2016 10:40 AM
End Date
13-5-2016 11:05 AM
Included in
Can Neuromorphic Computer Vision Inform Vision Science? Disparity Estimation as a Case Study
The primate visual system efficiently and effectively solves a multitude of tasks from orientation detection to motion detection. The Computer Vision community is therefore beginning to implement algorithms that mimic the processing hierarchies present in the primate visual system in the hope of achieving flexible and robust artificial vision systems. Here, we reappropriate the neuroscience “borrowed” by the Computer Vision community and ask whether neuromorphic computer vision solutions may give us insight into the functioning of the primate visual system. Specifically, we implement a neuromorphic algorithm for disparity estimation and compare its performance against that of human observers. The algorithm greatly outperforms human subjects when tuned with parameters to compete with non-neural approaches to disparity estimation on benchmarking stereo image datasets. Conversely, when the algorithm is implemented with biologically plausible receptive field sizes, spatial selectivity, phase tuning, and neural noise, its performance is directly relatable to that of human observers. The receptive field size and the number of spatial scales sensibly determine the range of spatial frequencies in which the algorithm successfully operates. The algorithm’s phase tuning and neural noise in turn determine the algorithm’s peak disparity sensitivity. When included, retino-cortical mapping strongly degrades disparity estimation in the model’s periphery, further closening human and algorithm performance. Hence, a neuromorphic computer vision algorithm can be reappropriated to model human behavior, and can provide interesting insights into which aspects of human visual perception have been or are yet to be explained by vision science.