The DOI matches its Crossref record; 143 of the 156 references it cites were found in a registry or an index; 10 of the 19 sources its claims cite were checked for support.
Large-Scale Unsupervised Semantic Segmentation
Open accessAbstract
Empowered by large datasets, e.g., ImageNet and MS COCO, unsupervised learning on large-scale data has enabled significant advances for classification tasks. However, whether the large-scale unsupervised semantic segmentation can be achieved remains unknown. There are two major challenges: i) we need a large-scale benchmark for assessing algorithms; ii) we need to develop methods to simultaneously learn category and shape representation in an unsupervised manner. In this work, we propose a new problem of large-scale unsupervised semantic segmentation (LUSS) with a newly created benchmark dataset to help the research progress. Building on the ImageNet dataset, we propose the ImageNet-S dataset with 1.2 million training images and 50k high-quality semantic segmentation annotations for evaluation. Our benchmark has a high data diversity and a clear task objective. We also present a simple yet effective method that works surprisingly well for LUSS. In addition, we benchmark related un/weakly/fully supervised methods accordingly, identifying the challenges and possible directions of LUSS. The benchmark and source code is publicly available at https://github.com/LUSSeg.
Overview
CoChat Summary
TL;DR
The paper proposes large-scale unsupervised semantic segmentation (LUSS) and introduces ImageNet-S, a benchmark built from ImageNet with 919 categories and nearly 50k pixel-level annotated images. It also presents a simple LUSS method using enhanced representation learning and pixel-attention, and benchmarks related un/weakly/fully supervised approaches.
Key findings
- ImageNet-S contains 1,183,322 training, 12,419 validation, and 27,423 testing images from 919 categories; 39,842 validation/testing images and 9,190 training images are annotated with precise pixel-level masks.
- The dataset includes ImageNet-S50 (50 categories; 64,431 train / 752 val / 1,682 test) and ImageNet-S300 (300 categories; 384,862 train / 4,097 val / 9,088 test) in addition to the full ImageNet-S.
- Most images contain one category, while 8.6% of images have more than one category.
- The authors correct 296 mislabeled images and find 942 images with missing labels during annotation.
Methods
- The paper defines LUSS as assigning labels to pixels from large-scale data without human-annotation supervision, requiring simultaneous learning of category and shape representations.
- The proposed method includes unsupervised representation learning with a non-contrastive pixel-to-pixel representation alignment strategy, a deep-to-shallow supervision strategy, and a pixel-attention scheme for label generation and fine-tuning.
- The benchmark is built from ImageNet by removing unsegmentable categories, annotating pixel-level masks, and evaluating related unsupervised, weakly supervised, and fully supervised methods.
Limitations
- The paper states that existing unsupervised semantic segmentation methods focus on small datasets and few easy categories, and some use supervised prior knowledge such as pre-trained weights, edge detection, or saliency detection, making evaluation difficult.
- It identifies LUSS challenges including simultaneous shape and category representation learning, semantic clustering at large scale, and high computational cost.
- The provided text is truncated before the experimental results and any explicit limitations section, so quantitative LUSS performance and author-stated limitations beyond the benchmark discussion are not available.
Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 2 Not valid · 2 Not checked · 18 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Existing segmentation datasets can hardly support LUSS due to the large image complexity but small data scale.”
§ 3 LARGE-SCALE UNSUPERVISED SEMANTIC SEG- > 3.1 Large-scale LUSS Dataset: ImageNet-S
What the sources say
- Partly supportedcites [10] sourceExperimental — accuracy not yet measured
Section 2.3.3 says the segmentation dataset needed to grow and that segmentation annotations required substantially more effort than detection annotations. Section 4.2 describes difficult images with occlusion and background clutter, but in an analysis of classification results. The paper does not say that existing segmentation datasets can hardly support LUSS or attribute such a limitation to the combination of image complexity and small data scale. (LLM, conf=0.80)
- Partly supportedcites [7] sourceExperimental — accuracy not yet measured
The Introduction states that existing urban-scene datasets are often small and do not fully capture the complexity of real-world inner-city traffic scenes, limiting progress in street-scene understanding. It does not extend that assessment to all segmentation datasets or specifically address LUSS. (LLM, conf=0.84)
- Partly supportedcites [8] sourceExperimental — accuracy not yet measured
The Introduction (p. 2) describes limitations in existing segmentation datasets’ object and scene coverage, and Table 1 (p. 8) compares their image counts and objects per image. These passages support concerns about dataset coverage and scale, but the paper does not assess whether those datasets can support LUSS or attribute any such difficulty to the combination of image complexity and small data scale. (LLM, conf=0.80)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“However, recent research [112], [113] suggested that performance on the ImageNet and downstream datasets is not always consistent due to the inconstancy of data distribution, data domain, and task objective.”
§ 3 LARGE-SCALE UNSUPERVISED SEMANTIC SEG- > 3.1 Large-scale LUSS Dataset: ImageNet-S
What the sources say
- Supportedcites [112] source
The ‘Pretraining domain matters’ section reports that in-domain pretraining consistently outperforms cross-domain pretraining, including ImageNet pretraining on iNat21. ‘SSL beyond ImageNet’ links differing results to pretraining and downstream data distributions, while §4.4 (‘Task granularity’) shows that performance also depends on the downstream classification task’s granularity. The copy provides section labels but no page numbers for these passages. (LLM, conf=0.87)
- Not checked — no full text of the sourcecites [113] source
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The training set and validation/testing sets have similar distributions.”
§ 3 LARGE-SCALE UNSUPERVISED SEMANTIC SEG- > 3.1 Large-scale LUSS Dataset: ImageNet-S
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“As shown in Table 1, the ImageNet-S dataset (see Figure 2) is much larger than existing datasets in terms of image amount and category diversity (see Figure 3).”
§ 3 LARGE-SCALE UNSUPERVISED SEMANTIC SEG- > 3.1 Large-scale LUSS Dataset: ImageNet-S
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Building on the ImageNet dataset, we propose the ImageNet-S dataset with 1.2 million training images and 50k high-quality semantic segmentation annotations for evaluation.”
- Also stated in: § 1 INTRODUCTION · hedged
“We remove the unsegmentable categories, e.g., bookshop, and utilize 919 categories with about 1.2 million images in ImageNet for training.”
- Also stated in: § 3.1 Large-scale LUSS Dataset: ImageNet-S · plain
“after removing the unsegmentable categories in the ImageNet dataset, e.g., bookshop, valley, and library, the ImageNet-S dataset contains 1,183,322 training, 12,419 validation, and 27,423 testing images from 919 categories.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1 INTRODUCTION · hedged
2 more claims
Create a free account to see all 26 claims and every source verdict.
- It may state two separate claims as one. 15
- Probably not a substantial claim. 6
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
U-Net: Convolutional Networks for Biomedical Image Segmentation
Ronneberger, O., Fischer, P., & Brox, T. (2015) · Lecture notes in computer science · doi:10.1007/978-3-319-24574-4_28 · Cited by 99,404
Densely Connected Convolutional Networks
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.243 · Cited by 48,566
Microsoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380
Fully convolutional networks for semantic segmentation
Long, J., Shelhamer, E., & Darrell, T. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298965 · Cited by 38,834
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00986 · Cited by 35,534
Rethinking the Inception Architecture for Computer Vision
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138
He, K., Gkioxari, G., Dollar, P., & Girshick, R. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.322 · Cited by 29,137
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.74 · Cited by 24,513
Research works printed but not linked: 4
- 10.5555/3524938.3525087 · doi:10.5555/3524938.3525087
- 10.5555/3495724.3497510 · doi:10.5555/3495724.3497510
- Russell, Segmenting scenes by matching image composites, 1580
- Li, Prototypical contrastive learning of unsupervised representations
Cited by
A Highly Efficient Model to Study the Semantics of Salient Object Detection
Cheng, M., Gao, S., Borji, A., Tan, Y., Zheng, L. L., & Wang, M. (2021) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3107956 · Cited by 112
Self-citationCorrMatch: Label Propagation via Correlation Matching for Semi-Supervised Semantic Segmentation
Sun, B., Yang, Y., Zhang, L., Cheng, M., & Hou, Q. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.00299 · Cited by 123
Self-citationAlpha-CLIP: A CLIP Model Focusing on Wherever you Want
Sun, Z., Fang, Y., Wu, T., Zhang, P., Zang, Y., Kong, S., Xiong, Y., Lin, D., & Wang, J. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.01237 · Cited by 156
Self-Supervised Learning of Object Parts for Semantic Segmentation
Ziegler, A., & Asano, Y. M. (2022) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52688.2022.01410 · Cited by 114
Masked Diffusion Transformer is a Strong Image Synthesizer
Gao, S., Zhou, P., Cheng, M.-M., & Yan, S. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.02117 · Cited by 91
Self-citation
Metadata
- Authors
- Gao, S., Li, Z.-Y., Yang, M.-H., Cheng, M.-M., Han, J., & Torr, P.
- Affiliations
- Nankai University (CN)
- University of California, Merced (US)
- Northwestern Polytechnical University (CN)
- University of Oxford (GB)
- Venue
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2023-06-01
- Volume
- 45
- Issue
- 6
- Pages
- 7457-7476
- ISSN
- 0162-8828 (ISSN-L)
- 2160-9292
- 1939-3539
- OpenAlex
- W4307844773
- Cited by
- 88
- References
- 156
- Citations by year
- 2026: 10 · 2025: 33 · 2024: 26 · 2023: 14
- FWCI
- 6.30
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Key Research and Development Program of China (2018AAA0100400)
- National Natural Science Foundation of China (62225604)
- Fundamental Research Funds for the Central Universities (63223050)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
- Keywords
- unsupervised semantic segmentation, benchmark dataset, weakly supervised segmentation, ImageNet
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Large-Scale Unsupervised Semantic Segmentation. https://app.cochat.ai/papers/10.1109/tpami.2022.3218275
- The paper
- Gao, S., Li, Z.-Y., Yang, M.-H., Cheng, M.-M., Han, J., & Torr, P. (2023). Large-Scale Unsupervised Semantic Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/tpami.2022.3218275
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.