Border Ownership, Category Selectivity and Beyond

Abstract

AbstractObject segmentation is a fundamental problem for both biological and machine vision systems. Recent advances in deep learning have allowed significant progress in terms of the ability of machine vision systems to carry out object segmentation, but this work has ignored a key piece of information that the biological vision system uses for segmentation: border ownership, or the determination for a given edge of which side the object is that owns it. Here we present a method for determining border ownership using a deep neural network model. Additionally, the model learns selectivity for object categories, suggesting a potential relationship between border ownership information and object category-selectivity. Our model may serve as a basic building block for machine vision systems that aim to reproduce the robustness of biological vision systems.KeywordsBorder ownershipCategory selectivityChannel representationT-junctionDeep neural networkObject segmentationTop-down guidanceOpposite channel ruleBorder ownership channelCategory (selective) channelLayer-0Global context awarenessDepth perception