A Software Retina for Egocentric Robotic Vision Applications on Mobile Platforms

摘要

We present work in progress to develop a low-cost highlyudintegrated camera sensor for egocentric and robotic vision. Our underlyingudapproach is to address current limitations to image analysis by DeepudConvolutional Neural Networks, such as the requirement to learn simpleudscale and rotation transformations, which contribute to the large computationaluddemands for training and opaqueness of the learned structure,udby applying structural constraints based on known properties of the humanudvisual system. We propose to apply a version of the retino-corticaludtransform to reduce the dimensionality of the input image space by audfactor of ex100, and map this spatially to transform rotations and scaleudchanges into spatial shifts. By reducing the input image size accordingly,udand therefore learning requirements, we aim to develop compact andudlightweight egocentric and robot vision sensor using a smartphone as theudtarget platform