Pith. sign in

REVIEW 2 major objections 8 minor 36 references

Scribble annotations recover 96% of full-mask segmentation for 2D materials

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 18:22 UTC pith:GUP7GBXW

load-bearing objection Scribble-supervised method for 2D material segmentation — solid engineering, but ARCL boundary loss appears to use full masks, contradicting the <0.6% annotation claim the 2 major comments →

arxiv 2607.07169 v1 pith:GUP7GBXW submitted 2026-07-08 cs.CV

TACoS: Weakly Supervised Learning of Two-Dimensional Materials from Scribble Annotations to Precise Segmentation

Jiabei Chen , Liping Zhang , Jiang-Bin Wu , Zhongming Wei , Enhao Ning , Su Yan , Weijun Li , Ping-Heng Tan
show 1 more author
Xin Ning
This is my paper
classification cs.CV
keywords scribble supervisionsemantic segmentationtwo-dimensional materialsweakly supervised learningcontrastive learningoptical microscopyminimum spanning treeconsistency regularization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces TACoS, a single-stage framework for pixel-level segmentation of two-dimensional material flakes from optical micrographs using only scribble annotations covering less than 0.6% of image pixels. The approach couples three mechanisms: weak-strong consistency alignment to propagate predictions across unlabeled regions, tree-energy regularization built from minimum spanning trees on backbone features to enforce structural coherence, and asymmetric regional contrastive learning to sharpen boundary discrimination in low-contrast edges. On graphene and MoS2 datasets, the method reaches over 96% of fully supervised mIoU while maintaining superior boundary IoU compared to existing weakly supervised baselines, including multi-stage approaches. The central claim is that these three components working jointly in a single training process can recover nearly all the performance of dense pixel-level annotation at a fraction of the labeling cost, specifically for the domain of 2D material microscopy where boundaries are governed by thin-film interference contrast and are often indistinct against complex substrates.

Core claim

Three jointly optimized constraints, consistency alignment on unlabeled pixels, minimum-spanning-tree-based soft pseudo-labels, and asymmetric contrastive learning on boundary regions, together recover over 96% of fully supervised segmentation performance from under 0.6% scribble annotation coverage on 2D material micrographs. The tree-energy module captures pixel affinity from both shallow color-gradient features and deep semantic features, generating online soft references that suppress foreground fragmentation. The contrastive module deliberately targets the hardest boundary pixels, those where augmented labels indicate foreground but strong-augmentation confidence collapses, applying rep

What carries the argument

TACoS combines: (1) Unlabeled Weak-Strong Distribution Alignment (UWSD), which uses cosine distance between weak and strong augmentation branch predictions on unlabeled pixels with stop-gradient on the weak branch; (2) Tree Energy Regularization (TER), which constructs minimum spanning trees on both shallow (Layer 2) and deep (Layer 11) DINOv2 features, computes pixel affinity from tree path distances, and generates structure-aware soft pseudo-labels via tree filtering; (3) Asymmetric Regional Contrastive Learning (ARCL), which fuses weak-branch high-confidence predictions with scribbles into augmented labels, mines hard samples where strong-branch confidence drops below threshold, computes类

Load-bearing premise

The weak-augmentation branch predictions used as pseudo-labels are assumed reliable enough to guide the strong branch, but if those predictions are systematically biased near low-contrast boundaries, the consistency, tree regularization, and contrastive modules could propagate that bias rather than correct it.

What would settle it

If TACoS is trained on a 2D material system where optical contrast between flake and substrate is substantially lower than in the tested graphene and MoS2 samples, the weak-branch pseudo-labels near boundaries would be more frequently wrong, and the 96%-of-full-supervision claim would likely break down.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Materials science labs could adopt automated flake screening without expert pixel-level annotation, reducing the bottleneck between sample exfoliation and characterization.
  • The tree-energy plus contrastive boundary design may transfer to other microscopy domains where object boundaries are governed by physics-based contrast rather than sharp edges, such as biological thin sections or crystallographic imaging.
  • If the 96%-of-full-supervision threshold holds across additional 2D material systems beyond graphene and MoS2, the scribble paradigm could become the default annotation standard for materials segmentation datasets.
  • The asymmetric boundary sampling strategy could be adapted for active learning, using the hard-sample set to suggest where additional scribble strokes would most improve the model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reliance on weak-branch pseudo-labels with a fixed confidence threshold of 0.7 creates a potential failure mode where systematically biased predictions near low-contrast boundaries are reinforced rather than corrected, since all three modules treat these pseudo-labels as ground truth. A confidence-adaptive or uncertainty-aware threshold might be necessary for materials with even lower optical
  • The single-stage design avoids multi-stage pseudo-label propagation but may sacrifice the error-correction opportunities that iterative refinement provides. Whether the tree-energy regularization can fully compensate for this in domains with more severe class imbalance remains an open question.
  • The method's boundary IoU, while improved over baselines, still lags substantially behind the mIoU gap to full supervision, suggesting that boundary precision remains the primary bottleneck even with the contrastive learning module.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 8 minor

Summary. The manuscript proposes TACoS, a single-stage scribble-supervised semantic segmentation framework for two-dimensional (2D) material flakes in optical microscopic images. The method integrates three components: (1) Unlabeled Weak-Strong Distribution alignment (UWSD), which enforces cosine consistency between weakly and strongly augmented views on unlabeled pixels; (2) Tree Energy Regularization (TER), which constructs minimum spanning trees on shallow and deep DINOv2 features to generate structure-aware soft pseudo labels; and (3) Asymmetric Regional Contrastive Learning (ARCL), which performs region-level prototype contrast on hard samples and applies an asymmetric boundary penalty. Experiments on graphene, MoS2, WSe2, and MoS2 datasets from two independent sources (Yan and Uslu) show that TACoS narrows the gap to fully supervised performance under sparse scribble annotations, with ablation studies isolating the contribution of each module.

Significance. The paper addresses a genuine practical bottleneck: pixel-level annotation of 2D material micrographs is expert-dependent and costly, limiting scalability of automated flake identification. The application of scribble supervision to this domain is novel, and the experimental design—two independent datasets, unified DINOv2 backbone for all baselines, boundary IoU as a supplementary metric, and module-level ablations—is reasonable. The framework is end-to-end trainable without iterative pseudo-label propagation, which is a practical advantage. However, the central quantitative claim of achieving over 96% of fully supervised performance using less than 0.6% annotated data is compromised by a supervision inconsistency in the ARCL module (see Major Comment 1), which must be resolved before the contribution can be properly assessed.

major comments (2)
  1. §3.2.3, Eqs. (15)–(16): The ARCL boundary loss L_Boundary operates on the geometric boundary set B, which is extracted from a full mask y_B. The text states: 'In the implementation, a full mask is uniformly applied to extract boundaries.' The weight w(i) in Eq. (16) is computed via cross-entropy against y_B(i), the full-mask boundary label. This means the ARCL module—which contributes +1.69% mIoU on graphene and +2.27% on MoS2 over UWSD+TER alone (Table 4)—directly uses dense pixel-level annotations during training. This contradicts the paper's core premise of scribble supervision and the claim of using less than 0.6% annotated data. The scribble annotations are generated from existing full masks (§4.1), so full masks are available, but using them for boundary supervision means the method is not purely scribble-supervised. The authors must either (a) replace the full-mask boundary labels
  2. Tables 2–4: No error bars, confidence intervals, or significance tests are reported for any experiment. Given that some improvements over baselines are modest (e.g., +1.74% over URSS on graphene, +1.91% over CC4S on MoS2), it is unclear whether these gains are statistically significant or within run-to-run variance. At minimum, the authors should report results over multiple random seeds (≥3) with mean and standard deviation for the main comparison table and the ablation study.
minor comments (8)
  1. Abstract and §4.1: The phrase 'less than 0.6% annotated data' is used prominently but the actual scribble coverage varies from 0.34% to 0.56% depending on dataset (Table 1). The abstract should specify which dataset this figure corresponds to, or use a range.
  2. §3.2.3, Eq. (11): The augmented label construction uses a fixed threshold τ_w=0.7. No sensitivity analysis for this threshold is provided. Given its role in pseudo-label quality, a brief study would strengthen the paper.
  3. §4.2: The training uses a single-GPU batch size of 2 across two GPUs (effective batch size 4). This is small for DINOv2 fine-tuning. The authors should comment on whether this affects convergence or stability.
  4. Figure 2: The overview diagram is dense and the text describing data flow between UWSD, TER, and ARCL is complex. A clearer separation of the three modules in the figure, with explicit arrow labels, would aid readability.
  5. §4.5, Figure 6: The failure case analysis is qualitative only. Quantifying the correlation between optical contrast ratio and misclassification rate would make this section more rigorous.
  6. §2.2: The related work discusses limitations of TEL's minimum spanning tree under low discriminative power of low-level features, but TER uses a similar MST construction. The authors should clarify how TER's dual-tree (shallow+deep) design specifically addresses the limitation they identify in TEL.
  7. Table 2: The 'Multi-stage' column is marked with ✓ for some methods but the criteria for this designation are not stated. A footnote defining what constitutes multi-stage would help.
  8. §3.2.4, Eq. (18): The loss weights λ_c=5, λ_t=1, λ_r=0.002, λ_b=0.00025 span several orders of magnitude. The justification given (inverse scaling between global and local gradients) is plausible but informal. A brief note on how these were tuned and whether they transfer to new materials would help.

Circularity Check

0 steps flagged

No formal circularity found; the framework's losses are defined against external scribble annotations and model-internal predictions, with results validated against external fully supervised baselines.

full rationale

The paper's derivation chain is not circular in the formal sense targeted by this analysis. The three loss components (UWSD, TER, ARCL) are defined against (a) external scribble annotations Ω_L, (b) model-internal predictions from the weak-augmentation branch used as stop-gradient pseudo-labels, and (c) feature-space representations of the current network state. None of these reduce to their own inputs by construction. The '96% of fully supervised performance' claim is verified against an external fully supervised DPT baseline (Table 2: 84.09% vs 86.33% on graphene, 85.20% vs 88.35% on MoS2), not against a fitted constant. The hyperparameters (λ_c, λ_t, λ_r, λ_b, τ_w, τ_s) are selected via grid search on validation sets (Section 4.6, Figure 7), not defined in terms of the target result. The skeptic's concern about ARCL's boundary loss using full-mask annotations (Section 3.2.3: 'a full mask is uniformly applied to extract boundaries') is a legitimate methodological inconsistency that inflates the practical annotation cost beyond the claimed <0.6%, but this is a correctness/consistency risk rather than a formal circularity: the boundary loss L_Boundary (Eq. 15) operates on geometric boundary sets extracted from masks, not on a quantity that is definitionally equivalent to the paper's claimed output. The self-citations (e.g., referencing Yan [23] for the dataset) are not load-bearing for the mathematical derivation. The framework's central claim has independent empirical content validated against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The paper introduces no new entities or particles. It uses standard neural network components (DINOv2, DPT decoder, minimum spanning trees, contrastive prototypes). The free parameters are hyperparameters tuned on validation data, with per-material optimization indicating some overfitting risk. The axioms are standard domain assumptions for weakly supervised learning, not ad hoc constructions.

free parameters (5)
  • σ_aff (TER affinity scale) = 0.1
    Set to empirical value of 0.1 for the Gaussian kernel bandwidth in the minimum spanning tree affinity matrix (Eq. 8, Section 4.2).
  • τ_w (weak confidence threshold) = 0.7
    Fixed threshold for pseudo-label filtering from weak augmentation branch (Eq. 11, Section 4.2).
  • τ_s (strong confidence threshold) = 0.97
    Fixed threshold for difficult sample mining in contrastive learning (Eq. 13, Section 4.2).
  • T (temperature coefficient) = 0.5
    Temperature for region contrastive loss softmax (Eq. 14, Section 4.2).
  • λ_c, λ_t, λ_r, λ_b (loss weights) = 5, 1, 0.002, 0.00025 (graphene); 5, 1, 0.0025, 0.00025 (MoS2)
    Loss balancing weights selected via grid search on validation set (Section 4.6, Figure 7). Different optimal values per material domain indicate dataset-specific tuning.
axioms (3)
  • domain assumption Weak augmentation predictions are reliable enough to serve as pseudo-labels for strong augmentation branch training.
    UWSD (Eq. 5) and ARCL (Eq. 11) both depend on weak branch predictions being approximately correct. This is assumed, not proven, and is the primary failure mode acknowledged in Section 4.5.
  • domain assumption Minimum spanning trees built on DINOv2 features capture meaningful pixel affinity relationships for 2D material boundaries.
    TER (Section 3.2.2) constructs MSTs on Layer 2 and Layer 11 features. The assumption is that feature distance correlates with semantic boundary structure, which may fail under low optical contrast.
  • domain assumption Scribble annotations provide sufficient shape priors to recover full segmentation masks.
    The entire framework assumes scribbles encode enough structural information. This is supported empirically by results but is a domain-specific assumption about 2D material morphology.

pith-pipeline@v1.1.0-glm · 21781 in / 2810 out tokens · 312918 ms · 2026-07-09T18:22:15.899220+00:00 · methodology

0 comments
read the original abstract

The precise pixel-level localization of 2D material flakes is crucial for high-throughput screening. However, traditional fully supervised methods rely on dense annotations, which are costly and time-consuming, severely limiting the practical deployment of segmentation models. This paper proposes TACoS, a specialized scribble segmentation framework tailored for 2D materials. First, we design a unified framework that integrates semi-supervised consistency learning with structured tree energy constraints. This framework comprises two core components: an unlabeled weak-strong distribution alignment module and a tree energy regularization module. The former employs cosine consistency constraints to enhance prediction alignment across views. Meanwhile, the latter utilizes minimum spanning trees to establish pixel affinity relationships and generate structure-aware soft pseudo labels for online semantic guidance. Next, we introduce asymmetric regional contrast learning. This approach fuses high-confidence predictions from the weak augmentation branch with scribbles to form augmented labels, and construct category prototypes in the representation space. Simultaneously, we prioritize contrastive constraints on challenging pixels in boundary-unlabeled regions. This strategy enhances intra-class cohesion and inter-class separation at the representation level, effectively reducing category confusion in low-contrast edges and complex backgrounds. Experiments conducted on the constructed graphene and MoS2 datasets demonstrate that our method TACoS achieves over 96% of fully supervised performance using less than 0.6% annotated data. Furthermore, it exhibits superior structural coherence and boundary stability in scenarios with weakly contrasting edges and complex backgrounds, providing an efficient and scalable solution for automated high-throughput screening of 2D material flakes.

Figures

Figures reproduced from arXiv: 2607.07169 by Enhao Ning, Jiabei Chen, Jiang-Bin Wu, Liping Zhang, Ping-Heng Tan, Su Yan, Weijun Li, Xin Ning, Zhongming Wei.

Figure 1
Figure 1. Figure 1: Annotation types illustration. We visualize four common annotation regimes for flake segmen￾tation: point, box, scribble, and full mask. Point and box annotations are placed on the foreground flake region, while scribble provides sparse strokes along the object structure; the full mask serves as the dense supervision reference. Existing scribble-supervised segmentation methods can be roughly categorized in… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed TACoS framework. The top layer performs weak–strong consistency learning with a shared DINOv2–DPT segmentation network, where weak predictions provide stop-gradient pseudo supervision while scribbles supply supervised signals. The bottom layer introduces two complemen￾tary structural modules. TER models pixel relations via a minimum spanning tree in the encoder feature space and ge… view at source ↗
Figure 3
Figure 3. Figure 3: Illustrative examples of the sparse structural supervision used in our framework. For each row, the left column shows the original microscopic image. To intuitively display the spatial distribution of the sparse supervision, the scribbles are directly superimposed onto the full masks in the right column. These scribbles capture key geometric structures while avoiding dense annotation, enabling effective le… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of flake segmentation on the Yan dataset. Visual comparison of segmen￾tation performance for graphene and MoS2 flakes. From top to bottom, the rows display the input optical microscopic images, ground truth (GT) masks, and prediction results from the baseline model, the fully su￾pervised model (Full), comparative methods (URSS for graphene and CC4S for MoS2), and our proposed TACoS fram… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of flake segmentation on the Uslu dataset. Visual comparison of segmenta￾tion results on graphene and WSe2 flakes from the Uslu dataset. The rows represent the input microscopic images, ground truth (GT) annotations, and predictions from the baseline model, the full-supervision model (Full), and our TACoS method. Misclassified regions are highlighted in red. TACoS consistently outputs r… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of segmentation failure cases under extreme optical conditions. The red pixels in the rightmost three columns highlight the misclassified regions (including both false positives and false negatives) produced by the fully supervised model (Full), the scribble-based baseline (Baseline), and the proposed TACoS method, respectively, when compared to the ground truth (GT). The examples are sampled… view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of the key hyper-parameters in TACoS. The line charts illustrate the fluc￾tuations in mIoU on the graphene and MoS2 datasets with respect to varying values of (a) λc, (b) λt , (c) λr , and (d) λb. an inversely proportional scaling mechanism. Specifically, the global loss is computed over a massive unlabeled grid, resulting in weak single-pixel gradients; conversely, the local loss gene… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 4 internal anchors

  1. [1]

    A. K. Geim, K. S. Novoselov, The rise of graphene, Nature Materials 6 (3) (2007) 183–191

    work page 2007

  2. [2]

    K. S. Novoselov, A. K. Geim, S. V . Morozov, D. Jiang, Y . Zhang, S. V . Dubonos, I. V . Grigorieva, A. A. Firsov, Electric field effect in atomically thin carbon films, Science 306 (5696) (2004) 666–669

    work page 2004

  3. [3]

    Blake, E

    P. Blake, E. W. Hill, A. H. Castro Neto, K. S. Novoselov, D. Jiang, R. Yang, T. J. Booth, A. K. Geim, Making graphene visible, Applied Physics Letters 91 (6) (Aug. 2007). 31

    work page 2007

  4. [4]

    X. Dong, H. Li, Y . Yan, H. Cheng, H. X. Zhang, Y . Zhang, T. D. Le, K. Wang, J. Dong, M. Jakobi, A. K. Yetisen, A. W. Koch, Deep-learning-based microscopic imagery classification, segmentation, and detection for the identification of 2D semiconductors, Advanced Theory and Simulations 5 (9) (2022) 2200140

    work page 2022

  5. [5]

    R. M. Sterbentz, K. L. Haley, J. O. Island, Universal image segmentation for optical identification of 2D materials, Scientific Reports 11 (1) (2021) 5808

    work page 2021

  6. [6]

    D. Lin, J. Dai, J. Jia, K. He, J. Sun, ScribbleSup: Scribble-supervised convo- lutional networks for semantic segmentation, in: Proceedings of the IEEE con- ference on computer vision and pattern recognition (CVPR), IEEE, 2016, pp. 3159–3167

    work page 2016

  7. [7]

    Y . B. Can, K. Chaitanya, B. Mustafa, L. M. Koch, E. Konukoglu, C. F. Baum- gartner, Learning to Segment Medical Images with Scribble-Supervision Alone, Springer International Publishing, 2018, pp. 236–244

    work page 2018

  8. [8]

    Z. Pan, P. Jiang, Y . Wang, C. Tu, A. G. Cohn, Scribble-supervised semantic seg- mentation by uncertainty reduction on neural representation and self-supervision on neural eigenspace, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2021, pp. 7396–7405

    work page 2021

  9. [9]

    Z. Pan, H. Sun, P. Jiang, G. Li, C. Tu, H. Ling, CC4S: Encouraging certainty and consistency in scribble-supervised semantic segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (12) (2024) 8918–8935

    work page 2024

  10. [10]

    Liang, T

    Z. Liang, T. Wang, X. Zhang, J. Sun, J. Shen, Tree Energy Loss: Towards sparsely annotated semantic segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), IEEE, 2022, pp. 16886–16895

    work page 2022

  11. [11]

    L. Wu, Z. Zhong, L. Fang, X. He, Q. Liu, J. Ma, H. Chen, Sparsely annotated semantic segmentation with adaptive gaussian mixtures, in: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), IEEE, 2023, pp. 15454–15464. 32

    work page 2023

  12. [12]

    L. Wu, Z. Zhong, J. Ma, Y . Wei, H. Chen, L. Fang, S. Li, Modeling the label distributions for weakly-supervised semantic segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (8) (2025) 6290–6306

    work page 2025

  13. [13]

    Zhang, J

    B. Zhang, J. Xiao, J. Jiao, Y . Wei, Y . Zhao, Affinity attention graph neural net- work for weakly supervised semantic segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 44 (11) (2022) 8082–8096

    work page 2022

  14. [14]

    H. Su, Y . Ye, W. Hua, L. Cheng, M. Song, SASFormer: Transformers for sparsely annotated semantic segmentation, in: Proceedings of the IEEE International Con- ference on Multimedia and Expo (ICME), IEEE, 2023, pp. 390–395

    work page 2023

  15. [15]

    DINOv2: Learning Robust Visual Features without Supervision

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fer- nandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, P. Bojanowski, DINOv2: Learn- ing robust visual features without s...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Ranftl, A

    R. Ranftl, A. Bochkovskiy, V . Koltun, Vision transformers for dense prediction, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2021, pp. 12159–12168

    work page 2021

  17. [17]

    Xiong, Z

    Y . Xiong, Z. Li, Y . Chen, F. Wang, X. Zhu, J. Luo, W. Wang, T. Lu, H. Li, Y . Qiao, L. Lu, J. Zhou, J. Dai, Efficient deformable ConvNets: Rethinking dynamic and sparse operator for vision applications, in: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), IEEE, 2024, pp. 5652–5661

    work page 2024

  18. [18]

    Masubuchi, E

    S. Masubuchi, E. Watanabe, Y . Seo, S. Okazaki, T. Sasagawa, K. Watanabe, T. Taniguchi, T. Machida, Deep-learning-based image segmentation integrated with optical microscopy for automatically searching for two-dimensional materi- als, npj 2D Materials and Applications 4 (1) (2020) 3

    work page 2020

  19. [19]

    J.-L. Uslu, T. Ouaj, D. Tebbe, A. Nekrasov, J. H. Bertram, M. Schütte, K. Watan- abe, T. Taniguchi, B. Beschoten, L. Waldecker, C. Stampfer, An open-source 33 robust machine learning platform for real-time detection and classification of 2D material flakes, Machine Learning: Science and Technology 5 (1) (2024) 015027

    work page 2024

  20. [20]

    H. Yang, R. Hu, H. Wu, X. He, Y . Zhou, Y . Xue, K. He, W. Hu, H. Chen, M. Gong, X. Zhang, P.-H. Tan, E. R. Hernández, Y . Xie, Identification and structural char- acterization of twisted atomically thin bilayer materials by deep learning, Nano Letters 24 (9) (2024) 2789–2797

    work page 2024

  21. [21]

    Zhang, H

    Y . Zhang, H. Zhang, S. Zhou, G. Liu, J. Zhu, Deep learning-based layer identifi- cation of 2D nanomaterials, Coatings 12 (10) (2022) 1551

    work page 2022

  22. [22]

    K. He, G. Gkioxari, P. Dollar, R. Girshick, Mask R-CNN, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2017, pp. 2980–2988

    work page 2017

  23. [23]

    S. Yan, J. Chen, X. Li, J. Bai, R. Zhou, L. Zhang, W. Li, J.-B. Wu, P.-H. Tan, X. Ning, Large scale and diverse two-dimensional flake segmentation dataset by general-purpose and labor-efficient annotation framework, npj 2D Materials and Applications 9 (1) (2025) 103

    work page 2025

  24. [24]

    Ouali, C

    Y . Ouali, C. Hudelot, M. Tami, Semi-supervised semantic segmentation with cross-consistency training, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2020, pp. 12671–12681

    work page 2020

  25. [25]

    X. Chen, Y . Yuan, G. Zeng, J. Wang, Semi-supervised semantic segmentation with cross pseudo supervision, in: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021, pp. 2613–2622

    work page 2021

  26. [26]

    Liang, X

    C. Liang, X. Zhao, J. Jia, J. Wang, L. Cao, J. Zhang, W. Chen, B 3CT: Three- branch learning with unlabeled target signals for domain-robust semantic seg- mentation, International Journal of Computer Vision 134 (3) (2026) 123

    work page 2026

  27. [27]

    W. Wang, T. Zhou, F. Yu, J. Dai, E. Konukoglu, L. V . Gool, Exploring cross-image pixel contrast for semantic segmentation, in: Proceedings of the IEEE/CVF inter- national conference on computer vision (ICCV), IEEE, 2021, pp. 7283–7293. 34

    work page 2021

  28. [28]

    Unsupervised Semantic Segmentation by Distilling Feature Correspondences

    M. Hamilton, Z. Zhang, B. Hariharan, N. Snavely, W. T. Freeman, Unsuper- vised semantic segmentation by distilling feature correspondences, arXiv preprint arXiv:2203.08414 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  29. [29]

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. Raffel, E. D. Cubuk, A. Kurakin, C.-L. Li, FixMatch: Simplifying semi-supervised learning with con- sistency and confidence, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 33, 2020, pp. 596–608

    work page 2020

  30. [30]

    L. Yang, Z. Zhao, H. Zhao, UniMatch V2: Pushing the limit of semi-supervised semantic segmentation, IEEE Transactions on Pattern Analysis and Machine In- telligence 47 (4) (2025) 3031–3048

    work page 2025

  31. [31]

    S. Liu, S. Zhi, E. Johns, A. Davison, Bootstrapping semantic segmentation with regional contrast, arXiv preprint arXiv:2104.04465 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  32. [32]

    S. Qiu, J. Chen, H. Zhang, R. Wan, X. Xue, J. Pu, Guided contrastive boundary learning for semantic segmentation, Pattern Recognition 155 (2024) 110723

    work page 2024

  33. [33]

    Boettcher, L

    W. Boettcher, L. Hoyer, O. Unal, J. E. Lenssen, B. Schiele, Scribbles for All: Benchmarking scribble supervised segmentation across datasets, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 37, 2024, pp. 46002– 46024

    work page 2024

  34. [34]

    Shrivastava, A

    A. Shrivastava, A. Gupta, R. B. Girshick, Training region-based object detectors with online hard example mining, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 761–769

    work page 2016

  35. [35]

    Decoupled Weight Decay Regularization

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [36]

    Cheng, R

    B. Cheng, R. Girshick, P. Dollár, A. C. Berg, A. Kirillov, Boundary IoU: Im- proving object-centric image segmentation evaluation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15334–15342. 35

    work page 2021