Information-Theoretic Segmentation by Inpainting Error Maximization

TL;DR

This paper introduces an unsupervised image segmentation method based on maximizing inpainting error, which is fast, class-agnostic, and effective even on single images, outperforming existing approaches.

Contribution

It presents a novel information-theoretic segmentation technique using inpainting error maximization that does not require deep training or labels.

Findings

Achieves state-of-the-art unsupervised segmentation quality.

Runs faster and is more general than competing methods.

Works effectively on individual unlabeled images.

Abstract

We study image segmentation from an information-theoretic perspective, proposing a novel adversarial method that performs unsupervised segmentation by partitioning images into maximally independent sets. More specifically, we group image pixels into foreground and background, with the goal of minimizing predictability of one set from the other. An easily computed loss drives a greedy search process to maximize inpainting error over these partitions. Our method does not involve training deep networks, is computationally cheap, class-agnostic, and even applicable in isolation to a single unlabeled image. Experiments demonstrate that it achieves a new state-of-the-art in unsupervised segmentation quality, while being substantially faster and more general than competing approaches.