Information-Theoretic Segmentation by Inpainting Error Maximization
This work provides a faster and more general unsupervised segmentation method for researchers and practitioners in computer vision, offering a class-agnostic solution applicable to single unlabeled images.
This paper addresses unsupervised image segmentation by partitioning images into maximally independent sets, aiming to minimize the predictability of one set from the other. The method achieves state-of-the-art unsupervised segmentation quality, outperforming competing approaches in speed and generality.
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.