Approaching human visual perception through AI-based representation of figure-ground segregation

TL;DR

This paper explores how AI models can mimic human visual perception by analyzing how borders are assigned to objects in complex visual scenes.

Contribution

The study introduces a novel approach to modeling border ownership inference using CNNs trained on occlusion stimuli and degraded contours.

Findings

Border ownership inference can emerge from feedforward computations even with degraded visual information.

Geometric context, especially junction-like configurations, significantly influences border ownership performance.

Hierarchical processing in CNNs leads to increasingly spatially coherent border ownership representations.

Abstract

Understanding how the visual system assigns borders to foreground objects is central to figure–ground perception, yet the computational principles underlying this process are still under investigation.

We trained multiple convolutional neural network (CNN) architectures on simple overlapping/occlusion stimuli and tested them on systematically degraded contours to probe how border-ownership (BOS) inference depends on available border context.

Across networks, BOS could be inferred from feedforward computations even under degraded conditions, but performance showed a strong dependence on junction-like configurations, indicating that geometric context contributes more than isolated edges. Accuracy increased approximately linearly with the amount of contextual information provided by fragmented borders, and representation analyses revealed a hierarchical progression from local edge responses to more spatially coherent, BOS-specific features.

Together, these results delineate which aspects of BOS can emerge from hierarchical feedforward processing and suggest that additional mechanisms such as horizontal and feedback interactions may reduce the visual information required for robust figure-ground segregation.

Figures