Do Neural Networks for Segmentation Understand Insideness?

TL;DR

This paper investigates whether neural networks can understand the insideness problem in image segmentation, revealing that recurrent networks trained on small images generalize well by effectively handling long-range spatial dependencies.

Contribution

It demonstrates that recurrent neural networks trained with small images can generalize to solve the insideness problem, highlighting their ability to evaluate long-range dependencies.

Findings

Few-unit DNNs can solve insideness for any curve.

Recurrent networks trained on small images generalize well.

Recurrent networks decompose long-range dependency evaluation.

Abstract

The insideness problem is an aspect of image segmentation that consists of determining which pixels are inside and outside a region. Deep Neural Networks (DNNs) excel in segmentation benchmarks, but it is unclear if they have the ability to solve the insideness problem as it requires evaluating long-range spatial dependencies. In this paper, the insideness problem is analysed in isolation, without texture or semantic cues, such that other aspects of segmentation do not interfere in the analysis. We demonstrate that DNNs for segmentation with few units have sufficient complexity to solve insideness for any curve. Yet, such DNNs have severe problems with learning general solutions. Only recurrent networks trained with small images learn solutions that generalize well to almost any curve. Recurrent networks can decompose the evaluation of long-range dependencies into a sequence of local operations, and learning with small images alleviates the common difficulties of training recurrent networks with a large number of unrolling steps.