Pith. sign in

REVIEW 4 major objections 5 minor 34 references

TextureSAM: Towards a Texture Aware Foundation Model for Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning SAM-2 on a texture-augmented version of ADE20K shifts its bias from object shape to surface texture, letting the resulting TextureSAM segment texture-defined regions more coherently than the original model.

desk verdict A plausible fine-tuning recipe for texture-aware SAM-2, but the headline gain rests on non-aggregated masks at an inference working point chosen after the fact; aggregated numbers don't support it. read the letter →

arxiv 2505.16540 v1 pith:66AD4RGR submitted 2025-05-22 cs.CV cs.AI

Inbal Cohen , Boaz Meivar , Peihan Tu , Shai Avidan , Gal Oren This is my paper · ORCID
classification cs.CVcs.AI
keywords texture-awaresegmentationshapebiasSegmentAnythingModelfine-tuningtextureaugmentationADE20Kcompositionalneuraltexturestexture-definedboundaries
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper takes a Segment Anything Model (SAM-2), which the authors argue is biased toward object shape because it was trained on semantic segmentation data, and fine-tunes it on a texture-augmented version of ADE20K. In that augmented dataset, object instances are dressed with textures from a separate texture database at varying strengths, so the model must learn to use surface patterns rather than shape priors. The result is TextureSAM, which the paper reports outperforms SAM-2 on two texture-defined benchmarks: RWTD natural images (0.47 vs. 0.26 mIoU without mask aggregation) and STMD synthetic images (0.35 vs. 0.07 mIoU), with higher ARI as well. The paper also shows that the fine-tuned model keeps most of its semantic segmentation ability on ADE20K. If true, this would mean a foundation segmenter can be shifted toward texture awareness by data augmentation alone, without changing the architecture.

What carries the argument

The load-bearing object is the Textured-ADE20K dataset, built by a compositional neural texture method. Each instance mask selects the Gaussians whose centers fall inside the mask; their appearance features are mixed with features sampled from a random texture as $\tilde{f}_c^m = \eta f_c^m + (1-\eta) f_t$, and the modified Gaussians are decoded into the augmented image. The coefficient $\eta$ controls how much texture replaces content, producing a training distribution that ranges from near-natural images to pure texture fields. Fine-tuning SAM-2 on this distribution is the mechanism that shifts the model's shape bias, and the two $\eta$ limits supply the paper's two variants.

What would settle it

Run the identical fine-tuning protocol on the original, un-texturized ADE20K with the same epochs, hyperparameters, and modified inference parameters, then compare its mIoU and ARI on RWTD and STMD with TextureSAM's; if the control reproduces the gains, the improvement is due to the inference working point rather than texture training.

Watch

Extended reading notes

Core claim

TextureSAM is a SAM-2 model fine-tuned on Textured-ADE20K, in which each ground-truth instance is filled with a randomly selected texture using a compositional neural texture transfer that interpolates content and texture features with a coefficient η. Training with mild augmentation (η≤0.3) yields the best natural-image texture segmentation, while strong augmentation (η≤1.0), where objects are fully replaced by texture, works best on the semantic-less synthetic dataset. The paper's central discovery is that this fine-tuning decouples shape from texture: the model stops fragmenting textured regions into semantic objects and instead predicts whole texture-defined areas, while retaining comparable semantic segmentation on ADE20K validation.

Load-bearing premise

The evaluation assumes that raising points-per-side from 32 to 64 and lowering the stability-score threshold from 0.95 to 0.2 is a neutral inference working-point change, not a tweak that favors the texture-trained model over the original.

Editorial extensions

If this is right

  • Without mask aggregation, TextureSAM raises mIoU on RWTD from 0.26 to 0.47 and on STMD from 0.07 to 0.35, with ARI rising from 0.36 to 0.62 and from 0.16 to 0.34.
  • The two augmentation strengths serve different regimes: strong texture replacement (η≤1.0) transfers best to the synthetic STMD benchmark, while mild augmentation (η≤0.3) preserves enough natural image statistics for the real-world RWTD benchmark.
  • Fine-tuning on ADE20K, a subset of SAM-2's original training data, limits catastrophic forgetting: on ADE20K validation, the mild variant's aggregated mIoU is 0.55, versus 0.46 for the original SAM-2 and 0.65 for SAM-2 evaluated with TextureSAM's inference parameters.
  • SAM-2's fragmentation of textured regions is a measurable failure mode: it produces many more predicted masks per ground-truth segment, and mask aggregation narrows or reverses the gap, implying that the texture-trained model's main advantage is coherent, unfragmented region prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper does not run: anneal η from mild to strong during training instead of using two fixed levels, which could produce a single model whose shape-texture trade-off is adjustable at inference.
  • If the protocol transfers, domains like histopathology and remote sensing, where boundaries are defined by surface patterns, could use the same augmentation recipe on any existing semantic dataset plus a texture library, without collecting new annotated masks.
  • Because the public release of Textured-ADE20K would let downstream users fine-tune on the η level matching their domain, the approach offers a practical knob for steering a foundation segmenter between shape-driven and texture-driven behavior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces TextureSAM, a SAM-2 model fine-tuned on a texture-augmented ADE20K dataset. Texture augmentation is performed with the Compositional Neural Textures method of [32], where instance regions are progressively replaced with textures from DTD under an interpolation coefficient η. The authors train two variants (mild η≤0.3 and strong η≤1.0) and evaluate them on the natural RWTD dataset and the synthetic STMD dataset, reporting mIoU and ARI both with and without mask aggregation, plus an ADE20K validation experiment. The central claim is that fine-tuning on texture-augmented data shifts SAM-2 from shape-biased to texture-aware segmentation, yielding large gains on texture-defined regions.

Significance. If the claim is supported, the paper would make a useful contribution: it proposes a concrete protocol for reducing shape bias in a segmentation foundation model, releases a texture-augmented version of ADE20K, and targets an application space (medical imaging, materials, remote sensing) where texture boundaries matter. The qualitative examples (Figures 1, 4, 6) do show reduced fragmentation of textured regions, which is a real and interesting effect. The central quantitative support, however, is entangled with a non-neutral change of inference parameters and with selective comparison baselines. The dataset and code release are explicitly promised, which is a strength, as are the dual metrics (mIoU and ARI) and the mask-aggregation analysis, which make the fragmentation issue visible. Yet, as written, the evaluation does not yet isolate the learned texture-awareness from inference working-point effects.

major comments (4)
  1. [§3.5, Tables 1 and 2] The modified inference parameters are not established as a neutral working point. Section 3.5 states that TextureSAM with default parameters produced no predicted masks for most images, so the authors changed points per side from 32 to 64 and stability score threshold from 0.95 to 0.2, and then evaluated all models at this point. Lowering the stability threshold by 0.75 and doubling the prompt density substantially changes the number, size, and confidence of predicted masks. Table 1 shows that SAM-2* with aggregation reaches 0.78 mIoU on STMD, above both TextureSAM variants (0.71 and 0.70); Table 2 shows SAM-2* with aggregation reaches 0.75 on RWTD, equal to TextureSAM η≤0.3 and just below η≤1.0. The claimed texture-awareness advantage is therefore only present without aggregation, at a working point selected because TextureSAM failed under defaults. The paper should either vary inference parameters systematically for each model (e.g., sweep stability threshold and points per side), report metrics at multiple working points, or match each model to its own optimal working point. Without this, the headline gains could be an artifact of threshold and mask-density effects rather than learned texture awareness.
  2. [Abstract, §4.2, Table 1] The abstract's '+0.18 mIoU' gain on STMD is not compared to the original SAM-2 but to SAM-2 with modified inference parameters. Table 1 reports original SAM-2 at 0.07 mIoU, SAM-2* at 0.17, and TextureSAM η≤1.0 at 0.35. The abstract says 'SAM-2', but the non-aggregated comparison that supports the '+0.18' figure appears to use SAM-2* as the baseline. The same ambiguity appears in Section 4.2, which states 'TextureSAM outperforms SAM-2' while the table shows SAM-2* in between. The claims should be matched to the exact baseline and experimental condition, and all four model/configurations should be reported consistently in the abstract and Section 4.
  3. [§4, Tables 1–3] No measure of variability is provided, so the significance of the reported differences cannot be assessed. The evaluation datasets are small (256 RWTD images; STMD is synthetic), the fine-tuning uses a single run with no mention of seeds, and no error bars or confidence intervals are reported. This matters directly for the main claim: for example, the non-aggregated RWTD mIoU gap between TextureSAM (0.47) and SAM-2* (0.14) is large, but the aggregated gap is zero or near zero, and the STMD aggregated gap reverses the conclusion. The authors should run multiple evaluation seeds (or bootstrap the datasets) and report error bars, at least for non-aggregated and aggregated mIoU and ARI.
  4. [§4.4, Table 3] The ADE20K validation experiment is used to claim that SEMANTIC segmentation capability is preserved, but the comparison is confounded by the same inference-parameter change. Table 3 shows SAM-2* at 0.65 aggregated mIoU, higher than TextureSAM η≤0.3 at 0.55, and the text acknowledges that 'modified inference parameters increased the original SAM-2's score by producing more masks.' This is exactly the kind of working-point effect that also drives the texture-dataset numbers. If the aggregation step plus denser masks is what produces the SAM-2* gains on ADE20K, then the paper needs to show that the texture gains on RWTD/STMD are not due to the same mechanism. At minimum, report the ADE20K results for each model at multiple inference parameter settings, including the default setting, and state explicitly how this inform the 'preserved semantic segmentation capability' claim.
minor comments (5)
  1. [§4.2] The sentence 'the η≤1.0 training includes images that are defined solely by texture boarders' contains a typo: 'boarders' should be 'boundaries'.
  2. [References] References [5] and [6] are the same DTD paper (one missing the third author, one with it), and [33] and [34] are the same ADE20K paper; the duplicates should be merged.
  3. [§3.4] The fine-tuning hyperparameters are said to be 'provided in the Supplemental Materials,' but no supplemental material is included in the submitted manuscript; either include the hyperparameters in the main text or cite the repository location clearly.
  4. [§3.3] The scaling step 'we scale the content image Ic into 8l×8l pixels' is ambiguous because l is not specified; please give the patch size and spacing in concrete pixel values.
  5. [§4.2, Figure 4] The box plot shows the number of predicted segments per number of GT segments, but the caption does not identify which model is which color; please add a legend or label each box.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity; minor self-citation in the texture-transfer backbone [32] and an inference-protocol confound, but the headline benchmark numbers are external.

  1. other [Section 1 and Section 3.3]
    "To do that, we rely on a recently introduced compositional neural texture approach (CNT) [32], that can interpolate between a source and style (i.e., texture) images. ... We employ the texture transfer technique in [32] to create the Textured-ADE20K dataset."

    The texture-transfer method used to build Textured-ADE20K is cited to prior work authored by co-author Peihan Tu [32]. This self-citation is load-bearing for constructing the fine-tuning data, so the method's core capability is imported rather than re-established. However, it is not a construction-level circularity: TextureSAM's mIoU/ARI numbers come from external held-out datasets (RWTD, STMD) and are not defined in terms of [32] or of any fitted parameter. The step is therefore a mild self-citation concern, not a reduction of the prediction to its input.

full rationale

The paper contains no derivation chain in which an output is fed back as an input. TextureSAM is produced by fine-tuning SAM-2 on a texture-augmented ADE20K set, and the headline gains are measured against external held-out texture benchmarks (RWTD and STMD) that are not used in training or in selecting the reported numbers. The only self-citation of note is [32] (CNT), shared with co-author Peihan Tu, which is used as the texture-transfer engine for constructing the training set; this is load-bearing for the method but not for the metric, so it does not make the conclusion equivalent to the citation. The inference working-point change in Section 3.5 is a real confound (default parameters yielded no masks for TextureSAM, so points_per_side and stability_score_thresh were moved to 64 and 0.2), but the same adjusted parameters were applied to SAM-2*, and the manuscript reports both aggregated and non-aggregated numbers; this is an evaluation-protocol concern, not a circular fit. Accordingly, the appropriate score is 2: a minor co-author self-citation with no construction-level circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central result rests on several domain assumptions about the texture transfer preserving boundaries, the benchmark datasets being valid, and the catastrophic-forgetting mitigation working. In addition, several hyperparameters (eta, epochs, inference parameters) are chosen by hand. These are not derived, so they carry experimental risk. No new physical or conceptual entities are postulated; TextureSAM is an engineered model variant.

free parameters (3)
  • texture interpolation coefficient eta thresholds = 0.3 (mild), 1.0 (strong)
    Chosen by hand to define mild versus strong texture augmentation; the choice changes which model wins on RWTD versus STMD and directly affects all reported comparisons.
  • inference parameters: points per side and stability score threshold = 64 and 0.2 (defaults were 32 and 0.95)
    Section 3.5 states the default inference parameters produced no predicted masks for most TextureSAM images, so these values were changed. All reported numbers therefore depend on this post-hoc working-point selection.
  • training epochs = 19 epochs for eta<=0.3, 25 for eta<=1.0
    Selected to balance training time with performance gains, rather than by a principled criterion; different epoch counts could change the comparisons.
assumptions (4)
  • domain assumption CNT texture transfer produces images whose ground-truth masks still describe the correct regions.
    Section 3.3 generates Textured-ADE20K by repainting mask regions with DTD textures; if the transfer degrades edges or blends across masks, the fine-tuning signal is corrupted.
  • domain assumption ADE20K training is a subset of SA-1B and therefore reduces catastrophic forgetting.
    Section 3.2 states this rationale for choosing ADE20K; it is plausible but not measured beyond the ADE20K validation table.
  • domain assumption RWTD and STMD ground truths capture texture-defined regions reliably.
    Section 3.2 describes RWTD and STMD as texture benchmarks; the entire evaluation depends on these annotations being correct, and no annotation-quality analysis is provided.
  • domain assumption SAM-2's strong semantic shape bias is the cause of fragmentation in texture scenarios.
    The conclusion states 'SAM-2 is indeed shape-biased'; this prior is reused from cited work and is consistent with the evidence but not proven by a controlled comparison that isolates shape from texture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TextureSAM: Towards a Texture Aware Foundation Model for Segmentation." pith.science (2026). https://pith.science/paper/66AD4RGR

@misc{pith2026250516540,
  author       = {Pith},
  title        = {Pith review of: TextureSAM: Towards a Texture Aware Foundation Model for Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66AD4RGR}},
  note         = {Machine review of arXiv:2505.16540}
}
read the original abstract

Segment Anything Models (SAM) have achieved remarkable success in object segmentation tasks across diverse datasets. However, these models are predominantly trained on large-scale semantic segmentation datasets, which introduce a bias toward object shape rather than texture cues in the image. This limitation is critical in domains such as medical imaging, material classification, and remote sensing, where texture changes define object boundaries. In this study, we investigate SAM's bias toward semantics over textures and introduce a new texture-aware foundation model, TextureSAM, which performs superior segmentation in texture-dominant scenarios. To achieve this, we employ a novel fine-tuning approach that incorporates texture augmentation techniques, incrementally modifying training images to emphasize texture features. By leveraging a novel texture-alternation of the ADE20K dataset, we guide TextureSAM to prioritize texture-defined regions, thereby mitigating the inherent shape bias present in the original SAM model. Our extensive experiments demonstrate that TextureSAM significantly outperforms SAM-2 on both natural (+0.2 mIoU) and synthetic (+0.18 mIoU) texture-based segmentation datasets. The code and texture-augmented dataset will be publicly available.

Figures

Figures reproduced from arXiv: 2505.16540 by the authors.

Figure 1
Figure 1. Examples for segmentation of natural images from the Real-World Textured Dataset. Compared with the original SAM (2nd and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Samples of Textured-ADE20K dataset. Incremental changes in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of generating textured image for dataset aug [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Box plot comparing predicted segments to the ground [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Segmentation results on images from the ADE20K [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Segmentation results on the synthetic STMD dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 26 canonical work pages

  1. [32]

    Compositional neural textures

    Peihan Tu, Li-Yi Wei, and Matthias Zwicker. Compositional neural textures. InSIGGRAPH Asia 2024 Conference Pa- pers, New York, NY , USA, 2024. Association for Computing Machinery

    work page 2024

  2. [1]

    Fowlkes, and Jitendra Malik

    Pablo Arbelaez, Michael Maire, Charless C. Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation.IEEE Trans. Pattern Anal. Mach. Intell., 33 (5):898–916, 2011

    work page 2011

  3. [2]

    Metal- lography and crystallographic texture analysis.The Encyclo- pedia of Archaeological Sciences, pages 1–4, 2018

    Arturo Barba-Pingarr ´on and Rafael Gonz ´alez-Parra. Metal- lography and crystallographic texture analysis.The Encyclo- pedia of Archaeological Sciences, pages 1–4, 2018

    work page 2018

  4. [3]

    Deepedge: A multi-scale bifurcated deep network for top- down contour detection

    Gedas Bertasius, Jianbo Shi, and Lorenzo Torresani. Deepedge: A multi-scale bifurcated deep network for top- down contour detection. In2015 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 4380– 4389, 2015

    work page 2015

  5. [4]

    Texture analysis of medical images.Clinical radiology, 59(12):1061–1069, 2004

    Gabriella Castellano, Leonardo Bonilha, LM Li, and Fer- nando Cendes. Texture analysis of medical images.Clinical radiology, 59(12):1061–1069, 2004

    work page 2004

  6. [5]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. InProceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014

    2014

  7. [6]

    De- scribing textures in the wild.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3613, 2014

    Mircea Cimpoi, Subhransu Maji, and Andrea Vedaldi. De- scribing textures in the wild.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3613, 2014

    work page 2014

  8. [7]

    Avoiding post-processing with context: Texture boundary detection in metallography

    Inbal Cohen, Julien Robitaille, Francis Quintal Lauzon, Ofer Beeri, Shai Avidan, and Gal Oren. Avoiding post-processing with context: Texture boundary detection in metallography. InAI for Accelerated Materials Design-NeurIPS 2024, 2024

    work page 2024

Show all 34 references
  1. [8]

    Mining textural knowl- edge in biological images: Applications, methods and trends

    Santa Di Cataldo and Elisa Ficarra. Mining textural knowl- edge in biological images: Applications, methods and trends. Computational and structural biotechnology journal, 15:56– 67, 2017

    2017

  2. [9]

    Piotr Doll ´ar and C. Zitnick. Fast edge detection using struc- tured forests.IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 2014

    2014

  3. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint ar...

    2010 arXiv

  4. [11]

    Robert M. French. Catastrophic interference in connectionist networks: Can it be predicted, can it be prevented? InAd- vances in Neural Information Processing Systems 6, pages 1176–1177, 1994

    1994

  5. [12]

    Are vision language models texture or shape biased and can we steer them?arXiv preprint arXiv:2403.09193, 2024

    Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Bianca Lamm, Muhammad Jehanzeb Mirza, Margret Keu- per, and Janis Keuper. Are vision language models texture or shape biased and can we steer them?arXiv preprint arXiv:2403.09193, 2024

    2024 arXiv

  6. [13]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. InInterna- tional conference on learning representations, 2018

    2018

  7. [14]

    Bdcn: Bi-directional cascade network for per- ceptual edge detection.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 44(1):100–113, 2022

    Jianzhong He, Shiliang Zhang, Ming Yang, Yanhu Shan, and Tiejun Huang. Bdcn: Bi-directional cascade network for per- ceptual edge detection.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 44(1):100–113, 2022

    2022

  8. [15]

    Segment anything model for medical images?arXiv preprint arXiv:2304.14660, 2023

    Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images?arXiv preprint arXiv:2304.14660, 2023

    2023 arXiv

  9. [16]

    Segment anything model for medical images?Medical Image Analysis, 92:103061, 2024

    Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images?Medical Image Analysis, 92:103061, 2024

    2024

  10. [17]

    Seg- ment anything is not always perfect: An investigation of sam on different real-world applications.arXiv preprint arXiv:2304.05750, 2023

    Wei Ji, Jingjing Li, Qi Bi, Wenbo Li, and Li Cheng. Seg- ment anything is not always perfect: An investigation of sam on different real-world applications.arXiv preprint arXiv:2304.05750, 2023

    2023 arXiv

  11. [18]

    Segment anything is not always perfect: An investi- gation of sam on different real-world applications, 2024

    Wei Ji, Jingjing Li, Qi Bi, Tingwei Liu, Wenbo Li, and Li Cheng. Segment anything is not always perfect: An investi- gation of sam on different real-world applications, 2024

    2024

  12. [19]

    Recent advances on image edge detec- tion: A comprehensive review.Neurocomputing, 2022

    Junfeng Jing, Shenjuan Liu, Gang Wang, Weichuan Zhang, and Changming Sun. Recent advances on image edge detec- tion: A comprehensive review.Neurocomputing, 2022

    2022

  13. [20]

    Learned shape-tailored descriptors for segmentation

    Naeemullah Khan and Ganesh Sundaramoorthi. Learned shape-tailored descriptors for segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5610–5619, 2018

    2018

  14. [21]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023

    2023

  15. [22]

    Pushing the boundaries of boundary de- tection using deep learning.arXiv: Computer Vision and Pattern Recognition, 2015

    Iasonas Kokkinos. Pushing the boundaries of boundary de- tection using deep learning.arXiv: Computer Vision and Pattern Recognition, 2015

    2015

  16. [23]

    Segment anything in medical images.Na- ture Communications, 15(1):44824, 2024

    Junde Ma, Yinan Chen, Yao Zhang, Ziqi Zhou, Xiaoping Yang, Yutong Xie, Yong Xia, Yi Xiong, Chunhua Shen, and Yongsheng Pan. Segment anything in medical images.Na- ture Communications, 15(1):44824, 2024

    2024

  17. [24]

    Martin, C.C

    D.R. Martin, C.C. Fowlkes, and J. Malik. Learning to detect natural image boundaries using local brightness, color, and 9 texture cues.IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(5):530–549, 2004

    2004

  18. [25]

    Have we solved edge detection? a review of state-of-the-art datasets and dnn based techniques.IEEE Access, 10:70541– 70555, 2022

    Muhammad Mubashar, Naeemullah Khan, Abdur Rehman Sajid, Muhammad Hashim Javed, and Naveed Ul Hassan. Have we solved edge detection? a review of state-of-the-art datasets and dnn based techniques.IEEE Access, 10:70541– 70555, 2022

    2022

  19. [26]

    Intriguing properties of vision transform- ers.Advances in Neural Information Processing Systems, 34: 23296–23308, 2021

    Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transform- ers.Advances in Neural Information Processing Systems, 34: 23296–23308, 2021

    2021

  20. [27]

    Edter: Edge detection with transformer

    Mengyang Pu, Yaping Huang, Yuming Liu, Qingji Guan, and Haibin Ling. Edter: Edge detection with transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1402–1412, 2022

    2022

  21. [28]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  22. [29]

    An end-to- end computer vision methodology for quantitative metallog- raphy.Scientific Reports, 12(1):4776, 2022

    Matan Rusanovsky, Ofer Beeri, and Gal Oren. An end-to- end computer vision methodology for quantitative metallog- raphy.Scientific Reports, 12(1):4776, 2022

    2022

  23. [30]

    Universal semantic-less texture boundary detection for mi- croscopy (and metallography)

    Matan Rusanovsky, Ofer Be’eri, Shai Avidan, and Gal Oren. Universal semantic-less texture boundary detection for mi- croscopy (and metallography). InMachine Learning and the Physical Sciences Workshop, NeurIPS, 2023

    2023

  24. [31]

    Deepcontour: A deep convolutional fea- ture learned by positive-sharing loss for contour detection

    Wei Shen, Xinggang Wang, Yan Wang, Xiang Bai, and Zhijiang Zhang. Deepcontour: A deep convolutional fea- ture learned by positive-sharing loss for contour detection. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3982–3991, 2015

    2015

  25. [33]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 633–641, 2017

    2017

  26. [34]

    Scene parsing through ade20k dataset.Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 633–641, 2017

    Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fi- dler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset.Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 633–641, 2017. 10

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.