Pith. sign in

REVIEW 2 major objections 5 minor 46 references

Towards Fine-grained Interactive Segmentation in Images and Videos

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read SAM2Refiner claims that freezing SAM2 and adding three lightweight modules—localization augment, prompt retargeting, and mask refinement—recovers fine-grained mask details in images and videos, outperforming prior state-of-the-art while…

desk verdict Image half is solid and well-ablated; video half is likely confounded by an unmatched per-frame prompting protocol, so the headline claim needs a re-run before it holds. read the letter →

arxiv 2502.09660 v1 pith:YQG34XY5 submitted 2025-02-12 cs.CV eess.IV

Yuan Yao , Qiushi Yang , Miaomiao Cui , Liefeng Bo This is my paper · ORCID
classification cs.CVeess.IV
keywords interactivesegmentationSegmentAnythingModelSAM2Refinerfine-grainedvideoobjectmaskrefinementpromptretargetinglocalizationaugment
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 claims that SAM2, the video-capable successor to SAM, can be turned into a high-precision interactive segmenter without retraining its backbone. The proposed SAM2Refiner freezes SAM2 and inserts three trainable components—localization augment, prompt retargeting, and mask refinement—so that the model keeps its global semantic understanding and streaming video memory while recovering fine boundaries and small structures. The payoff, if true, is that a single foundation model serves both everyday and detail-critical segmentation tasks, matching or beating dedicated high-quality segmenters. On four fine-grained image benchmarks the reported average mIoU is 93.1 versus 88.2 for HQ-SAM, and on the LVOS video benchmarks it reports J&F gains over both SAM2 and HQ-SAM2 without any video training.

What carries the argument

The load-bearing object is the SAM2Refiner architecture stacked on the frozen SAM2 image encoder and mask decoder. It has three named components: Localization Augment (LA), which splits the input into four sub-images, upsamples them, encodes them with the full image, pools local features at multiple granularities, and lets global features cross-attend to them; Prompt Retargeting (PR), which builds a dense prompt map from clicks and a coarse mask, encodes it with small convolutions, adds it to a multi-branch receptive-field-block-augmented object embedding, and renews the embedding through attention steps that reuse SAM2's sparse prompt and output tokens; and Mask Refinement (MR), a UNet-like decoder that upsamples and fuses the object embedding with multi-scale encoder features, ending in two transposed convolutions and a low-level feature addition to produce a 1024x1024 mask. Together they are inserted along SAM2's pipeline, and at video inference the same modules operate inside SAM2's streaming memory loop while re-using the first-frame prompt.

What would settle it

Re-run LVOS v1/v2 with SAM2 and HQ-SAM2 under the same per-frame first-prompt re-use protocol used for SAM2Refiner and compare J&F; if the gap shrinks to near zero, the video improvement is mostly prompting protocol, not the modules.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the three modules together let a frozen SAM2 produce fine-grained masks that surpass specialist high-quality segmentation models. The Localization Augment module crops the image into four overlapping sub-images, encodes them alongside the full image, and uses cross-attention so global features borrow local detail; the Prompt Retargeting module realigns dense prompt maps (clicks and an initial mask) with the augmented object embedding before renewing it through attention, restoring responsiveness to prompts; and the Mask Refinement module fuses the object embedding with multi-scale encoder features through a UNet-like decoder to output a 1024x1024 mask instead of SAM2's interpolated 256x256 prediction. Trained only on image data (HQSeg-44K) with the SAM2 backbone frozen, the method reports average mIoU/mBIoU of 93.1/88.9 on DIS, COIFT, HRSOD, and ThinObject-5K, compared to 88.2/80.3 for HQ-SAM, and J&F of 80.9 and 79.9 on LVOS v1/v2, compared to 77.2/78.3 for SAM2.

Load-bearing premise

The load-bearing premise is that the video evaluation protocol—re-using the first frame's prompt for every subsequent frame—was applied identically to SAM2Refiner and to the SAM2 and HQ-SAM2 baselines; if those baselines ran under standard streaming memory with only the initial prompt, the video gains could largely reflect per-frame prompting rather than the new modules.

Editorial extensions

If this is right

  • On the four fine-grained image benchmarks, average mIoU rises from 88.2 for HQ-SAM to 93.1, with the largest gains on thin and highly detailed objects.
  • Video fine-grained segmentation improves without any video training: LVOS v1 J&F goes from 77.2 for SAM2 and 79.6 for HQ-SAM2 to 80.9, and LVOS v2 from 78.3 and 79.1 to 79.9.
  • Point-prompt responsiveness increases sharply, especially with few clicks: with a single point, mIoU is 25.2 for SAM2Refiner versus 17.7 for HQ-SAM, and with ten points it reaches 93.1.
  • The design is plug-in: freezing SAM2 and training only the three modules preserves SAM2's existing zero-shot and streaming capabilities, so the same backbone can be deployed for general and fine-grained tasks.
  • The 1024x1024 mask output removes the interpolation step in SAM2's 256x256 decoder prediction, which is where boundary detail is lost.

Reading between the lines

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

  • Because the modules are trained only on images yet improve video results, they likely learn generic boundary and detail cues that transfer across modalities; one testable extension is applying the same three modules to SAM2 variants or other hierarchical encoders.
  • The video evaluation re-uses the first-frame prompt on every frame; if this per-frame prompting protocol was not applied identically to the SAM2 and HQ-SAM2 baselines, part of the reported video gap could be protocol-driven rather than module-driven.
  • The four-way crop in LA is a fixed spatial tiling; an adaptive version that crops around the prompt or predicted object could reduce compute or further improve detail recovery, but the paper does not test this.
  • Combining PR's dense prompt encoding with a diffusion-based refinement head is a natural next step, since the paper's MR module only uses convolutional fusion and may saturate on very thin structures.
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

2 major / 5 minor

Summary. The paper proposes SAM2Refiner, a fine-grained interactive segmentation framework built on a frozen SAM2 backbone. Three modules are introduced: Localization Augment (LA), which cross-attends global features with multi-granularity local features extracted from four upsampled sub-images; Prompt Retargeting (PR), which aligns a dense prompt map with the augmented object embedding; and Mask Refinement (MR), which fuses multi-scale encoder features with the object embedding in a UNet-like decoder to predict high-resolution masks. The model is trained on HQSeg-44K only and then evaluated zero-shot on four image benchmarks (DIS, COIFT, HRSOD, ThinObject-5K) and two video benchmarks (LVOS v1 and v2). The paper reports average mIoU/mBIoU of 93.1/88.9 with the Hiera-large backbone, outperforming HQ-SAM, and J&F of 80.9/79.9 on LVOS v1/v2, with an ablation study in Table 4 supporting the contribution of each module.

Significance. The image-side contribution is well supported internally: the comparisons to SAM and HQ-SAM use the same four benchmarks and the same HQSeg-44K training data, and the ablation in Table 4 shows a plausible attribution of the gains to the proposed modules. The architectural ideas—local-global cross attention for detail enrichment and spatially aligned prompt retargeting—are potentially reusable for other SAM-family models. If the video results survive a corrected evaluation protocol, the paper would be a useful contribution to fine-grained interactive segmentation. However, the video evidence is currently unreliable because of an asymmetric prompting protocol (Section 3.3 versus Section 4.1), and the strongest video baseline, HQ-SAM2, is undefined. The abstract's dual image-and-video claim therefore hinges on a re-evaluation of Table 2. The manuscript does not release code, which limits reproducibility, but the described experimental protocol is sufficiently detailed for the image part.

major comments (2)
  1. [Section 3.3 (Inference) and Table 2] The video evaluation uses a non-standard and potentially unfair protocol. Section 3.3 states that 'for each frame except the first frame, the model re-uses the prompt input from the first frame as the original prompt to perform frame-by-frame segmentation.' This gives SAM2Refiner the first-frame prompt on every frame, whereas the standard semi-supervised VOS protocol invoked in Section 4.1 ('semi-supervised setting with the first frame prompt given') supplies the prompt only on the first frame and then relies on memory propagation. The paper does not document whether the SAM2 and HQ-SAM2 baselines in Table 2 received the same per-frame reprompting. If they did not, the reported gains on LVOS v1 (J&F 77.2 for SAM2 vs 80.9 for SAM2Refiner) could be caused entirely by the extra prompt signal rather than by the three proposed modules. Please report (i) the exact prompt protocol used for every baseline in Table 2, and (ii) SAM2Refiner's results under the standard first-frame-only protocol.
  2. [Table 2] The baseline 'HQ-SAM2 [16]' is not defined anywhere in the manuscript. Reference [16] is the image-only HQ-SAM paper (Lei Ke et al., NeurIPS 2023), and no construction, training recipe, or weight source is given for a video-capable 'HQ-SAM2'. Because this is the strongest competing video method in Table 2, omitting its definition makes the video comparison unverifiable. Please specify how HQ-SAM2 was built (e.g., SAM2 memory combined with HQ-SAM's decoder), where its weights came from, and under which prompt protocol it was evaluated.
minor comments (5)
  1. [Section 4.2] The text says 'SAM2Refiner with a smaller backbone (hiera_l) performs 93.1% and 88.9% average mIoU and mBIoU', but hiera_l is the large backbone in Table 1; the smaller backbone is hiera_b+. Please correct this wording.
  2. [Table 4] The check-mark rows are ambiguous. After the SAM2 baseline row, the checkmarks do not clearly indicate which modules are active; the row beginning 'SAM2Refiner' with two checkmarks is immediately followed by another row with two checkmarks, so the reader cannot tell whether the final row (check, check, check) is the only full model. Please label each row explicitly (e.g., LA only, LA+PR, LA+MR, PR+MR, full model).
  3. [References] References [28] and [29] are duplicate entries for the same SAM2 arXiv preprint; Table 1 and Table 2 cite SAM2 as [28], while Table 4 cites it as [29]. Please consolidate to a single reference.
  4. [Section 3.2.2] The training procedure samples bounding boxes, points, and coarse masks, but the dense prompt map is described as encoding only clicks and masks (positive click, negative click, mask in three channels). The paper does not explain how bounding-box prompts are represented in D_m. Please clarify this for reproducibility.
  5. [Section 3.2.1] The LA module processes five full-resolution encoder passes per image (the original image plus four upsampled sub-images), but no runtime, GPU memory, or parameter count is reported. Since the method targets interactive use, the computational overhead should be quantified.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the image and video claims are benchmark measurements against external baselines, with only a background self-citation ([27]) that is not load-bearing.

full rationale

This is an empirical refinement paper, not a derivation. SAM2Refiner freezes SAM2 and adds LA/PR/MR; the paper reports benchmark results (Tables 1-4) and an ablation in which modules are removed one at a time. No reported metric is obtained by fitting a parameter to the target value and then rereading it as a prediction; no uniqueness theorem or ansatz is imported from the authors' prior work, and the only author self-citation ([27], Qiushi Yang) appears in a list of traditional CNN-based segmentation methods and is not load-bearing. The image results, including the 93.1 mIoU/88.9 mBIoU average and the point-prompt study, are evaluated under a protocol shared with the baselines and therefore are self-contained evidence. A separate concern is the video protocol: Sec. 3.3 states that 'for each frame except the first frame, the model re-uses the prompt input from the first frame as the original prompt to perform frame-by-frame segmentation,' and Table 2 does not state whether the SAM2 and HQ-SAM2 baselines received the same per-frame re-prompting. This is a potential measurement asymmetry for the video comparisons, but it is not a self-referential reduction of a derivation to its own inputs, so it does not constitute circularity under the stated criteria. Score 1 reflects the minor background self-citation and the video-protocol uncertainty; no claim in the paper is circular by construction.

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

The ledger reflects that the method is an empirical architecture paper: hyperparameters are hand-selected with no sensitivity analysis, the key domain assumptions are that frozen SAM2 features and HQSeg-44K supervision transfer to the test benchmarks, and that the video evaluation protocol is fair. No new physical or mathematical entities are introduced.

free parameters (4)
  • intermediate loss weight lambda_i = 0.3
    Chosen by hand to weight intermediate supervision in Eq. 3; no sensitivity study is provided, and the final result depends on this value.
  • LA pooling receptive fields = 2, 4, 8
    The three multi-granularity pooling scales in Eq. 1 are hand-selected with n=3; the LA ablation shows sensitivity to the module, but the specific scales are not varied.
  • RFB branch kernel sizes = 3, 5, 7
    Asymmetric convolution kernel widths in the prompt retargeting module are set by hand to give multi-size receptive fields.
  • prompt map resolution = 256 x 256
    The dense prompt map is fixed at 256x256 to lower computational cost, affecting the alignment granularity for prompt features.
assumptions (4)
  • domain assumption SAM2's frozen encoder and decoder features are a strong enough base that only the new modules need training
    Stated in Sec. 3.2 and 3.3 ('we fix the model parameters of the pre-trained SAM2 model'); if false, the method's zero-shot gains would not transfer.
  • domain assumption HQSeg-44K provides sufficient fine-grained supervision to generalize zero-shot to DIS validation, ThinObject test, COIFT and HR-SOD
    Training set in Sec. 3.3 includes DIS and ThinObject-5K training splits; the open-world generalization claim depends on this distribution transfer.
  • ad hoc to paper Upsampling four cropped sub-images to full resolution does not introduce harmful noise and preserves local details
    The LA module in Sec. 3.2.1 assumes that 2x zoomed crops, upsampled back, provide useful local context; no artifact analysis is provided.
  • ad hoc to paper Feeding the first-frame prompt into every video frame is a fair evaluation protocol and matches how baselines were run
    Stated in Sec. 3.3 for inference; the paper does not show baselines under the same protocol, making video comparisons potentially inconsistent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Fine-grained Interactive Segmentation in Images and Videos." pith.science (2026). https://pith.science/paper/YQG34XY5

@misc{pith2026250209660,
  author       = {Pith},
  title        = {Pith review of: Towards Fine-grained Interactive Segmentation in Images and Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YQG34XY5}},
  note         = {Machine review of arXiv:2502.09660}
}
read the original abstract

The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer performance degradations in scenarios demanding accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between the ability to perceive intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. To this end, we present an SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns and maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embedding, we introduce a prompt retargeting module to renew the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2502.09660 by the authors.

Figure 1
Figure 1. Comparison of segmentation results of the proposed framework SAM2Refiner and previous methods. The proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed framework SAM2Refiner. It contains a localization augment module to balance the detailed and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. A Comparison between SAM2 and SAM2Refiner. SAM2Refiner introduces three modules for fine-grained interactive segmen [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with previous methods. Given the blue box as visual prompt, our proposed SAM2Refiner produces more [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Effectiveness of LA module. LA suggest remarkable [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Effectiveness of PR module. PR allows precise response [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 32 canonical work pages

  1. [13]

    Rethink- ing space-time networks with improved memory coverage for efficient video object segmentation

    Yu-Wing Tai Ho Kei Cheng and Chi-Keung Tang. Rethink- ing space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021

    work page 2021

  2. [29]

    Sam 2: Segment anything in images and videos

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

    arXiv 2024

  3. [16]

    Segment anything in high quality

    Martin Danelljan Yifan Liu Yu-Wing Tai Chi-Keung Tang Fisher Yu Lei Ke, Mingqiao Ye. Segment anything in high quality. In NeurIPS, 2023

    work page 2023

  4. [1]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

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

    arXiv 2010

  5. [2]

    Per-pixel classification is not all you need for semantic segmentation

    Alexander Kirillov Bowen Cheng, Alex Schwing. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021

    work page 2021

  6. [3]

    Wang Y . Chen J. Jiang D. Zhang X. Tian Q. Cao, H. and Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, 2022

    work page 2022

  7. [4]

    Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model

    Ho Kei Cheng and Alexander G Schwing. Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022

    work page 2022

  8. [5]

    CascadePSP: Toward class-agnostic and very high- resolution segmentation via global and local refinement

    Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. CascadePSP: Toward class-agnostic and very high- resolution segmentation via global and local refinement. In CVPR, 2020

    work page 2020

Show all 46 references
  1. [6]

    Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar

    Ishan Misra Alexander G. Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022

    2022

  2. [7]

    Collins Yukun Zhu Ting Liu Thomas S

    Maxwell D. Collins Yukun Zhu Ting Liu Thomas S. Huang- Hartwig Adam Cheng, Bowen and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020

    2020

  3. [8]

    Global contrast based salient region detection

    Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. TPAMI, 37(3):569–582, 2014

    2014

  4. [9]

    Instance-aware se- mantic segmentation via multi-task network cascades

    Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware se- mantic segmentation via multi-task network cascades. In CVPR, 2016

    2016

  5. [10]

    Seg- ment and recognize anything at any granularity

    Peize Sun Xueyan Zou Shilong Liu Chunyuan Li-Jianwei Yang Lei Zhang Jianfeng Gao Feng Li, Hao Zhang. Seg- ment and recognize anything at any granularity. In ECCV, 2024

    2024

  6. [11]

    Lvis: A dataset for large vocabulary instance segmentation

    Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. InCVPR, 2019

    2019

  7. [12]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Gir- shick. Mask r-cnn. In ICCV, 2017

    2017

  8. [14]

    Panoptic segmentation

    Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollár. Panoptic segmentation. In CVPR, 2019

    2019

  9. [15]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. In ICCV, 2023

    2023

  10. [17]

    Fss-1000: A 1000-class dataset for few- shot segmentation

    Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few- shot segmentation. In CVPR, 2020

    2020

  11. [18]

    Rethinking atrous convolution for semantic image segmentation

    Florian Schroff Hartwig Adam Liang-Chieh Chen, George Papandreou. Rethinking atrous convolution for semantic image segmentation. In arXiv preprint arXiv:1706.05587, 2017

    2017 arXiv

  12. [19]

    Deep interactive thin object selection

    Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Ji- ashi Feng. Deep interactive thin object selection. In WACV, 2021

    2021

  13. [20]

    Segment anything with precise interaction

    Mengzhen Liu, Mengyu Wang, Henghui Ding, Yilong Xu, Yao Zhao, and Yunchao Wei. Segment anything with precise interaction. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3790–3799, 2024

    2024

  14. [21]

    Receptive field block net for accurate and fast object detection

    Songtao Liu, Di Huang, et al. Receptive field block net for accurate and fast object detection. In Proceedings of the Eu- ropean conference on computer vision (ECCV) , pages 385– 400, 2018

    2018

  15. [22]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015

    2015

  16. [23]

    Segment anything in medical images

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024

    2024

  17. [24]

    Gaus- sian grouping: Segment and edit anything in 3d scenes

    Fisher Yu Mingqiao Ye, Martin Danelljan and Lei Ke. Gaus- sian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024

    2024

  18. [25]

    Recurrent dynamic embedding for video object segmentation

    Zhiwei Xiong Bang Zhang Pan Pan Mingxing Li, Li- ucheng Hu and Dong Liu. Recurrent dynamic embedding for video object segmentation. In CVPR, 2022

    2022

  19. [26]

    Highly accurate dichotomous im- age segmentation

    Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous im- age segmentation. In ECCV, 2022

    2022

  20. [27]

    Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation

    Yixuan Yuan Qiushi Yang. Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation. In MIC- CAI Workshop, 2020

    2020

  21. [30]

    Hi- era: A hierarchical vision transformer without the bells-and- whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hi- era: A hierarchical vision transformer without the bells-and- whistles. In International Conference on Machine Learn...

    2023

  22. [31]

    Anything- 3d: Towards single-view anything reconstruction in the wild

    Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything- 3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023

    2023 arXiv

  23. [32]

    Hierarchical image saliency detection on extended cssd

    Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. TPAMI, 38(4): 717–729, 2015

    2015

  24. [33]

    Segrefiner: Towards model- agnostic segmentation refinement with discrete diffusion process

    Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model- agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023

    2023 arXiv

  25. [34]

    Sun K. Cheng T. Jiang B. Deng C. Zhao Y .-Liu D. Mu Y . Tan M. Wang X. Wang, J. and W. Liu. Deep high-resolution rep- resentation learning for visual recognition. In TPAMI, 2020

    2020

  26. [35]

    Seggpt: Segmenting ev- erything in context

    Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting ev- erything in context. arXiv preprint arXiv:2304.03284, 2023

    2023 arXiv

  27. [36]

    Medical sam adapter: Adapting seg- ment anything model for medical image segmentation.arXiv preprint arXiv:2304.12620, 2023

    Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting seg- ment anything model for medical image segmentation.arXiv preprint arXiv:2304.12620, 2023

    2023 arXiv

  28. [37]

    Edit every- thing: A text-guided generative system for images editing

    Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit every- thing: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023

    2023 arXiv

  29. [38]

    Al- varez Xie, Enze and Ping Luo

    Wenhai Wang Zhiding Yu Anima Anandkumar Jose M. Al- varez Xie, Enze and Ping Luo. Segformer: Simple and effi- cient design for semantic segmentation with transformers. In NeurIPS, 2021

    2021

  30. [39]

    Embodiedsam: Online segment any 3d thing in real time

    Xiuwei Xu, Huangxing Chen, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Embodiedsam: Online segment any 3d thing in real time. arXiv preprint arXiv:2408.11811, 2024

    2024 arXiv

  31. [40]

    Seg- ment everything everywhere all at once

    Hao Zhang Feng Li Linjie Li Jianfeng Wang-Lijuan Wang Jianfeng Gao Yong Jae Lee Xueyan Zou, Jianwei Yang. Seg- ment everything everywhere all at once. In NeurIPS, 2023

    2023

  32. [41]

    Saliency detection via graph-based man- ifold ranking

    Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based man- ifold ranking. In CVPR, 2013

    2013

  33. [42]

    Decoupling features in hi- erarchical propagation for video object segmentation

    Zongxin Yang and Yi Yang. Decoupling features in hi- erarchical propagation for video object segmentation. In NeurIPS, 2022

    2022

  34. [43]

    Multi-view aggregation network for dichoto- mous image segmentation

    Qian Yu, Xiaoqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichoto- mous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3921–3930, 2024

    2024

  35. [44]

    Inpaint anything: Segment anything meets image inpainting

    Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023

    2023 arXiv

  36. [45]

    Segfix: Model-agnostic boundary refinement for segmenta- tion

    Yuhui Yuan, Jingyi Xie, Xilin Chen, and Jingdong Wang. Segfix: Model-agnostic boundary refinement for segmenta- tion. In ECCV, 2020

    2020

  37. [46]

    Towards high-resolution salient object detec- tion

    Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detec- tion. In ICCV, 2019

    2019

  38. [47]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017

    2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.