Autoregressive Unsupervised Image Segmentation

Ouali, Yassine; Hudelot, Céline; Tami, Myriam

doi:10.1007/978-3-030-58571-6_9

Cited by 71 publications

(77 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Ouali et al . [27] did a follow-up work on IIC. In addition to standard image transformations, they proposed to process image patches through various masked convolutions.…”

Section: Related Workmentioning

confidence: 99%

“…In this section, we first review how recent work [17,27] uses MI for unsupervised semantic image segmentation. Then, we discuss limitations of these methods, and how we tackle them in InfoSeg.…”

Section: Motivationmentioning

confidence: 99%

“…State-of-the-art methods [17,27] adapt the MI based image clustering approach of IIC [17] for segmentation. In the following, we introduce IICs' approach on image clustering and then the proposed modifications for segmentation.…”

Section: Unsupervised Semantic Image Segmentationmentioning

confidence: 99%

“…We motivate the exact structure of InfoSeg by first giving a thorough review of current-state-of-the-art methods [17,27], followed by a discussion of their limitations and how we approach them in InfoSeg. Our qualitative and quantitative evaluation show that InfoSeg significantly outperforms all compared methods.…”

Section: Introductionmentioning

confidence: 99%

“…For example, we achieve a relative increase of 26% in Pixel-Accuracy (PA) on the COCO-Stuff dataset [4]. Even though we follow the standard evaluation protocol for quantitative evaluation, we provide a critical discussion of it and uncover problems left undiscussed by recent work [17,27]. Furthermore, in addition to established datasets, we introduce COCO-Persons as a novel benchmark.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

InfoSeg: Unsupervised Semantic Image Segmentation with Mutual Information Maximization

Harb¹,

Knöbelreiter²

2021

Preprint

View full text Add to dashboard Cite

We propose a novel method for unsupervised semantic image segmentation based on mutual information maximization between local and global high-level image features. The core idea of our work is to leverage recent progress in self-supervised image representation learning. Representation learning methods compute a single high-level feature capturing an entire image. In contrast, we compute multiple high-level features, each capturing image segments of one particular semantic class. To this end, we propose a novel two-step learning procedure comprising a segmentation and a mutual information maximization step. In the first step, we segment images based on local and global features. In the second step, we maximize the mutual information between local features and high-level features of their respective class. For training, we provide solely unlabeled images and start from random network initialization. For quantitative and qualitative evaluation, we use established benchmarks, and COCO-Persons, whereby we introduce the latter in this paper as a challenging novel benchmark. InfoSeg significantly outperforms the current state-of-the-art, e.g., we achieve a relative increase of 26% in the Pixel Accuracy metric on the COCO-Stuff dataset.

show abstract

“…Ouali et al . [27] did a follow-up work on IIC. In addition to standard image transformations, they proposed to process image patches through various masked convolutions.…”

Section: Related Workmentioning

confidence: 99%

Section: Motivationmentioning

confidence: 99%