Pith. sign in

REVIEW 4 major objections 5 minor 51 references

U-CFR: Uncertainty-Guided Cascade Forward Refinement for Interactive Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read U-CFR claims that an interactive segmenter can correct itself by generating pseudo-clicks at uncertain boundaries, cutting required clicks by over 10% on Berkeley.

desk verdict The EDH results are real, but the uncertainty-guided pseudo-click mechanism that gives the paper its title looks self-contradictory and the CFR gains are within noise. read the letter →

arxiv 2607.20705 v1 pith:SYQLK3HO submitted 2026-07-22 cs.CV cs.AI

Elijah Danquah Darko , Min Xian , Terence Soule , Tiankai Yao , Matthew William Anderson This is my paper · ORCID
classification cs.CVcs.AI
keywords interactivesegmentationuncertainty-guidedrefinementpseudo-clicksedgedetectioncascadeforwardclickefficiencyboundary-awareuncertaintyout-of-domaingeneralization
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

U-CFR (Uncertainty-Guided Cascade Forward Refinement) claims that an interactive segmentation model can correct its own mask between user interactions by placing internal pseudo-clicks on the boundary regions where it is most uncertain. The paper couples a dual-head network—one head for region masks, one for edge maps—with an inference-time refinement loop driven by the map Ubd = Upred · Gseg, the product of predictive uncertainty and the contour-gradient magnitude. If the claim is right, annotation becomes cheaper: fewer human clicks reach the same mask quality, and the improvement shows up both on natural-image benchmarks and on out-of-domain medical and material-science datasets. The paper reports, for instance, that the average number of clicks needed to reach 90% IoU on Berkeley falls from 2.46 to 2.19, and that the one-click IoU on OAIZIB rises from 2.52% to 6.32%.

What carries the argument

The load-bearing machinery is the boundary-aware uncertainty map Ubd = Upred · Gseg, where Upred = 1 - |2Pseg - 1| quantifies predictive uncertainty and Gseg is the normalized gradient magnitude of the segmentation probability map (computed with a Sobel filter). Because both factors must be high, Ubd concentrates pseudo-clicks on predicted boundaries that the model is unsure about. The second piece is the selective labeling rule (Eq. 7), which assigns a positive or negative pseudo-click only when Pseg is above 0.51 or below 0.49, and skips the ambiguous interval in between; the pseudo-click is then appended to the user click set and re-fed through the model, producing the cascade-forward ref

What would settle it

Compute the ground-truth agreement rate of the pseudo-clicks: for each image, check whether the positive/negative label generated by Eq. 7 at argmax Ubd matches the true object label, and correlate that agreement with the measured NoC reduction; if most pseudo-clicks fall in the skipped interval or reinforce confidently wrong predictions, the uncertainty mechanism is not the source of the gain.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that boundary-aware uncertainty can drive a self-correction loop for interactive segmentation. After each user click, the model produces a probability map Pseg; it then forms Upred = 1 - |2Pseg - 1|, an uncertainty map that peaks where the model is close to a coin flip, and Gseg, the normalized Sobel-gradient magnitude of Pseg. Their product Ubd = Upred · Gseg is high only at pixels that are simultaneously low-confidence and on a predicted contour. The model places a pseudo-click at the maximum of Ubd, labels it positive if Pseg > 0.51 or negative if Pseg < 0.49 (and skips it in the [0.49,0.51] interval), then feeds the expanded click set ba

Load-bearing premise

The paper depends on the assumption that the model's self-generated clicks are corrective: the pixels the map highlights are precisely those where the model is least certain, and the click label is the model's own guess, so if the model is confidently wrong the pseudo-click can reinforce the mistake.

Editorial extensions

If this is right

  • Fewer human clicks: the method reduces clicks to reach 90% IoU on Berkeley from 2.46 to 2.19 and improves NoC@90 on Pascal VOC from 2.81 to 2.64.
  • Better first impression: the edge-aware base model lifts one-click mask quality (mIoU@1) on COCO_MVal by about 8%, and on OAIZIB the one-click IoU rises from 2.52% to 6.32%.
  • Sharper boundaries: normalized surface distance scores are consistently as good or better than the baseline, especially at early clicks.
  • Out-of-domain transfer: without any fine-tuning, the method reduces clicks on unseen medical and material-science images, e.g., SEM-Precipitate NoC@85 drops from 10.37 to 7.92 relative to the closest compared baseline.
  • The refinement loop adds no trainable parameters: all improvements after training come from the inference-time pseudo-click cascade.

Reading between the lines

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

  • Because the pseudo-click rule needs only a probability map, the same Ubd recipe could in principle be applied to other click-based segmenters, not just the ViT-based dual-head model trained here; the paper does not test this transfer, but it is a direct consequence if the uncertainty signal is what carries the benefit.
  • The paper's own numbers suggest the uncertainty loop's marginal contribution over plain cascade refinement is small in some settings; a reader should be cautious about attributing the 10% headline gain to the pseudo-clicks rather than to the edge-aware training, unless an ablation holds the base model fixed and varies only the pseudo-click generation.
  • A testable modification would be to place the pseudo-click just off the argmax, at the nearest pixel with Pseg outside the [0.49,0.51] skip interval, since the current rule sometimes has no valid label at the most uncertain boundary point; this could reveal whether skipping ambiguous pixels helps or hurts.
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. This manuscript proposes U-CFR, an inference-time refinement framework for interactive segmentation. The method uses a ViT-B/MAE backbone with two heads: a segmentation head and an auxiliary edge-detection head (EDH), trained jointly with a weighted loss. At inference, U-CFR computes a boundary-aware uncertainty map Ubd = Upred·Gseg (Eq. 6), places a pseudo-click at its argmax, labels the click by thresholding the model's own probability Pseg via Eq. (7), and feeds the updated click set back into a cascade-forward refinement loop. The paper reports that EDH improves upon the SimpleClick baseline in click efficiency, initial mask quality, and boundary accuracy, and that U-CFR further reduces the number of clicks. It claims over 10% click reduction on Berkeley relative to SimpleClick and reports out-of-domain gains on medical and SEM datasets. Code is provided.

Significance. If the claims are valid, U-CFR would be a useful training-free, inference-time self-correction mechanism for interactive segmentation. The EDH ablation is well designed and shows consistent improvements over the SimpleClick baseline, and the out-of-domain evaluation is a strength of the paper. However, the central novelty—uncertainty-guided pseudo-clicks—is not convincingly established. The pseudo-click mechanism is internally inconsistent: the uncertainty map is maximal where Pseg≈0.5, which is exactly the interval where Eq. (7) instructs the method to skip. Moreover, the pseudo-click labels are derived from the model's own prediction, so they carry no external corrective information. The empirical gains over the CFR baseline in Table 2 are small and unreplicated, and the headline 'over 10%' claim comes from a comparison against SimpleClick rather than against the CFR ablation. The paper's main contribution therefore remains unsupported.

major comments (4)
  1. [§3.3, Eqs. (4)–(7)] Ubd is the product of Upred and Gseg. Upred is maximal when Pseg≈0.5, and Gseg is large on predicted boundaries. Thus the argmax of Ubd will often fall in the interval [0.49, 0.51] where Eq. (7) explicitly says 'skip'. No fallback is specified, so on those steps the U-CFR loop degenerates to plain CFR. Please report the fraction of pseudo-clicks that are actually placed, the distribution of Pseg at the argmax locations, and a concrete fallback rule. Without this, the proposed uncertainty-guided loop is not well defined.
  2. [§3.3, Eq. (7) and Table 2] The pseudo-click label is computed by thresholding Pseg, the model's own output. A click labeled 'positive' at a pixel where Pseg>0.51 or 'negative' where Pseg<0.49 agrees with the current prediction, so it cannot correct a confidently wrong mask; it can even reinforce a false positive or false negative. The only information that differs from plain CFR is the click location, which is also derived from Pseg. Please test the informativeness of the pseudo-clicks by ablating with (a) clicks at random boundary points, (b) flipped labels, and (c) a fixed uncertainty threshold that actually places clicks inside the skipped band. If the small gains over CFR persist in these ablations, they likely come from the CFR mask-guidance loop rather than from the uncertainty mechanism.
  3. [Table 2 and §4.3] The claimed U-CFR gains over CFR are 0.04–0.08 clicks in NoC@90 (2.72→2.64 on PascalVOC, 3.88→3.80 on COCO_MVal, 4.98→4.94 on SBD). No standard deviations, number of seeds, or significance tests are reported, and the differences are small enough to be within evaluation noise. Furthermore, NSDS@10 is identical (0.40, 0.73, 0.73) or slightly worse at NSDS@20 on COCO_MVal (0.82 vs 0.83), which does not support the abstract's boundary-accuracy claim for U-CFR. Please provide repeated-run statistics and a significance test, or temper the claims accordingly.
  4. [Abstract and Table 3] The 'over 10%' reduction on Berkeley compares Base+U-CFR1 (NoC@90=2.19) against SimpleClick (2.46), not against the CFR or EDH baselines. This number conflates the contributions of the dual-head architecture, the CFR loop, and the uncertainty-guided pseudo-clicks. The isolated U-CFR contribution in Table 2 is only 2–3%, which is within plausible noise. Please re-frame the headline and, if possible, provide a full decomposition (Base vs Base+EDH vs Base+EDH+CFR vs Base+EDH+U-CFR) with confidence intervals.
minor comments (5)
  1. [§4.1, Eq. (1)] The symbol α is used for two different quantities: first as the focal-loss balancing parameter in 'Normalized Focal Loss with α=0.5' and later as the segmentation-loss weight in L_total with 'α=0.2 and β=2'. Please rename one of them for clarity.
  2. [Eq. (3)] The formula for the normalized focal loss appears malformed: the denominator '|Pseg|' is undefined. Please state the normalization explicitly, including the summation domain.
  3. [Table 3] The text says 'new state-of-the-art across nearly all evaluated metrics', but on DAVIS the NoC@85 of Ours(Base+U-CFR1) is 4.29, worse than SimpleClick's 4.10. Please qualify the claim and mark the best values consistently.
  4. [§3.3] The thresholds τ−=0.49 and τ+=0.51 are said to be set 'from our experiments', but no ablation or selection procedure is given. A sensitivity analysis over these thresholds would help reproducibility, especially since they encode the skip behavior that is central to the method.
  5. [Figure 3] Figure 3 is not discussed in the text. Please add a reference and a brief explanation of what the refinement panels show.

Circularity Check

1 steps flagged · score 4.0 of 10

U-CFR's pseudo-clicks are generated from the model's own Pseg, so the claimed corrective signal is self-referential; the EDH comparison is independent, which limits the circularity.

  1. self definitional [Section 3.3, Eq. (4)-(7) (pseudo-click generation; claimed in Abstract and Section 4.3)]
    "Upred = 1 − |2·Pseg −1| ... Ubd = Upred · Gseg ... label(i*,j*) = skip, τ− ≤ Pseg(i*,j*) ≤ τ+; positive, Pseg(i*,j*) > τ+; negative, Pseg(i*,j*) < τ−. ... This rule deliberately skips placing a pseudo-click in the ambiguous interval between τ− and τ+, where the model's prediction is least certain and essentially noise."

    The pseudo-click location is argmax of Ubd, which is maximal exactly where Pseg≈0.5 by Eq. (4), i.e., inside the [0.49,0.51] interval that Eq. (7) marks 'skip'; when a click is placed, its label is the model's own thresholded Pseg. Thus the 'corrective signal' is constructed from the very prediction it is supposed to correct: a confident false-positive region receives a positive pseudo-click and a confident false-negative region receives a negative one, so no pseudo-click can oppose the current error. Any measured gain over CFR is therefore not attributable to the uncertainty mechanism as a corrective signal; the loop reduces to the CFR mask-guidance mechanism plus a self-confirming input from the same network.

full rationale

The EDH component is evaluated independently in Table 1 by comparing the dual-head model with the SimpleClick baseline under identical training except for the edge head, so that part of the paper is self-contained and not circular. The U-CFR-vs-CFR comparison in Table 2 is also an external empirical benchmark, and the paper itself concedes that 'the CFR baseline eventually achieves comparable accuracy after 10 clicks'; the measured gains are small (0.04-0.08 clicks). The circularity is confined to the interpretation that U-CFR's uncertainty-guided pseudo-clicks provide 'strong corrective signals': Equations (4)-(7) define both the click location and the click label from the same Pseg, so the pseudo-click cannot carry information independent of the model's current prediction. No load-bearing self-citation or imported uniqueness theorem is present; CFR-ICL is used as a baseline, not as a proof. Overall score 4: one central self-referential mechanism, but the framework's empirical backbone and EDH ablation remain independent enough that the paper is not wholly circular.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central mechanism is a tuned inference heuristic: two thresholds, two loss weights, click radius, and cascade length are chosen on experimental data; no sensitivity analysis is reported. The 'self-correction' label derives from the model's own Pseg, so the method's corrective signal is internally generated rather than anchored to an external error source.

free parameters (6)
  • τ− (lower confidence threshold) = 0.49
    Chosen 'from our experiments' (§3.3, Eq. 7) to classify/skip pseudo-clicks; no sensitivity analysis.
  • τ+ (upper confidence threshold) = 0.51
    Chosen 'from our experiments' (§3.3, Eq. 7); tiny interval around the Ubd maximum region.
  • α (segmentation loss weight in L_total) = 0.2 (and also used as focal-loss α=0.5 in the same section)
    Loss weights 'selected by experiments' (§4.1); notation conflict between focal-loss α and total-loss α.
  • β (edge loss weight) = 2
    Selected by experiments (§4.1) to emphasize edge supervision.
  • Number of U-CFR cascade steps = not reported; tables use CFR1/U-CFR1
    Method says 'a fixed number of steps' but no value is given; the one-step variant is the one evaluated.
  • Click disk radius = 5 pixels
    Implementation choice for encoding user clicks (§4.1).
assumptions (4)
  • domain assumption Supervising an edge head with Sobel-derived GT edges from segmentation masks improves boundary features without hurting segmentation
    Foundation of EDH design; evaluated only on final metrics, not separately verified.
  • ad hoc to paper The argmax of Ubd = Upred·Gseg identifies a useful pseudo-click site
    Introduced in §3.3 Eq. 6; no independent evidence that this map correlates with actual segmentation errors.
  • ad hoc to paper The model's own confidence Pseg at a candidate pixel is a reliable basis for labeling a corrective click
    Eq. 7 labels as positive/negative solely from Pseg; if the model is confidently wrong, the pseudo-click reinforces the error.
  • domain assumption The CFR loop with the same model and clicks is a valid control for U-CFR
    Used for Table 2 comparisons to attribute gains to the uncertainty mechanism.
invented entities (1)
  • Boundary-aware uncertainty map Ubd = Upred·Gseg
    purpose: Generates pseudo-click locations during inference
    A new composite signal defined only from the model's own output; no testable prediction outside the method's evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of U-CFR: Uncertainty-Guided Cascade Forward Refinement for Interactive Segmentation." pith.science (2026). https://pith.science/paper/SYQLK3HO

@misc{pith2026260720705,
  author       = {Pith},
  title        = {Pith review of: U-CFR: Uncertainty-Guided Cascade Forward Refinement for Interactive Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYQLK3HO}},
  note         = {Machine review of arXiv:2607.20705}
}
read the original abstract

Interactive image segmentation is critical for efficient image annotation; however, existing methods often require many corrective clicks or rely on passive refinement schemes that converge slowly. We propose Uncertainty-Guided Cascade Forward Refinement (U-CFR), a novel inference-time framework that enables models to autonomously self-correct after each user interaction. U-CFR introduces a boundary-aware uncertainty score that fuses segmentation uncertainty, contour gradients, and explicit edge predictions to guide the placement of internal pseudo-clicks. These self-generated clicks target the most ambiguous boundary regions, providing strong corrective signals without additional manual input. To support this process, we design a dual-head network with a shared encoder-decoder backbone: a segmentation head ensures region consistency, while an edge head sharpens boundary alignment. In inference, U-CFR launches a cascade of refinement steps, where each stage leverages the uncertainty-driven pseudo-clicks to refine the mask progressively. Experiments on standard benchmark datasets demonstrate that the proposed U-CFR improves click efficiency, initial mask quality, and boundary accuracy. It reduces the required clicks by over 10% on challenging datasets like Berkeley and offers a more intelligent and efficient interactive annotation.

Figures

Figures reproduced from arXiv: 2607.20705 by the authors.

Figure 1
Figure 1. Overview of the proposed Uncertainty-Guided Cascade Forward Refinement (U-CFR) framework. User [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison between our proposed U-CFR model and SimpleClick Liu et al. [2023]. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Sample of U-CFR refinement. The input image (with a positive click) is processed to generate a prediction [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 6 linked inside Pith

  1. [1]

    ICCV , @pages=

    Semantic contours from inverse detectors , author=. ICCV , @pages=. 2011 , @organization=

    2011

  2. [2]

    ECCV , pages=

    Microsoft COCO: Common objects in context , author=. ECCV , pages=. 2014 , @organization=

    2014

  3. [3]

    IJCV , volume=

    The pascal visual object classes (voc) challenge , author=. IJCV , volume=. 2010 , @publisher=

    2010

  4. [4]

    GrabCut

    " GrabCut" interactive foreground extraction using iterated graph cuts , author=. TOG , @volume=. 2004 , @publisher=

    2004

  5. [5]

    ICCV , @volume=

    A Database of Human Segmented Natural Image , author=. ICCV , @volume=. 2001 , @organization=

    2001

  6. [6]

    CVPR , @pages=

    A benchmark dataset and evaluation methodology for video object segmentation , author=. CVPR , @pages=

  7. [7]

    arXiv:2107.02314 , year=

    The RSNA-ASNR-MICCAI BraTS 2021 Benchmark , author=. arXiv:2107.02314 , year=

    arXiv 2021

  8. [8]

    figshare , @pages=

    Segmented anisotropic ssTEM dataset of neural tissue , author=. figshare , @pages=. 2013 , @publisher=

    2013

Show all 51 references
  1. [9]

    ICCV , @pages=

    Simpleclick , author=. ICCV , @pages=

  2. [10]

    arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv:2010.11929 , year=

    2010 arXiv

  3. [11]

    V-Net: Fully Convolutional Neural Networks for Volumetric Segmentation , year=

    Milletari, Fausto and Navab, Nassir and Ahmadi, Seyed-Ahmad , booktitle=. V-Net: Fully Convolutional Neural Networks for Volumetric Segmentation , year=

  4. [12]

    ICCV , @pages=

    Segment anything , author=. ICCV , @pages=

  5. [13]

    IEEE/CVF CVPR , @pages=

    Focalclick , author=. IEEE/CVF CVPR , @pages=

  6. [14]

    AAAI , @volume=

    CFR-ICL , author=. AAAI , @volume=

  7. [15]

    NeurIPS , @volume=

    Segment everything everywhere all at once , author=. NeurIPS , @volume=

  8. [16]

    WACV , @pages=

    Deep interactive thin object selection , author=. WACV , @pages=

  9. [17]

    PAMI , @volume=

    Small-object sensitive segmentation using across feature map attention , author=. PAMI , @volume=. 2022 , @publisher=

    2022

  10. [18]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Small-Object Sensitive Segmentation Using Across Feature Map Attention , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  11. [19]

    CVPR , @pages=

    Deep Residual Learning , author=. CVPR , @pages=

  12. [20]

    PAMI , @volume=

    Deep High-Resolution Representation Learning , author=. PAMI , @volume=. 2020 , publisher=

    2020

  13. [21]

    CVPR , @pages=

    Masked autoencoders are scalable vision learners , author=. CVPR , @pages=

  14. [22]

    arXiv:2410.12214 , year=

    Order-Aware Interactive Segmentation , author=. arXiv:2410.12214 , year=

    arXiv

  15. [23]

    TCSVT , year=

    MST: Adaptive Multi-Scale Tokens Guided Interactive Segmentation , author=. TCSVT , year=

  16. [24]

    CVPR , @pages=

    Cross-Domain Multi-Task Learning , author=. CVPR , @pages=

  17. [25]

    MICCAI , pages=

    Boundary-Aware Transformers , author=. MICCAI , pages=. 2021 , @organization=

    2021

  18. [26]

    arXiv:2312.12425 , year=

    SegRefiner: Towards model-agnostic segmentation refinement with discrete diffusion process , author=. arXiv:2312.12425 , year=

    arXiv

  19. [27]

    ECCV , @pages=

    Interactive image segmentation with click imitation , author=. ECCV , @pages=

  20. [28]

    NeurIPS , volume=

    Pytorch , author=. NeurIPS , volume=

  21. [29]

    CVPR , pages=

    FocusCut , author=. CVPR , pages=

  22. [30]

    CVPR , pages=

    Interactive image segmentation with latent diversity , author=. CVPR , pages=

  23. [31]

    CVPR , pages=

    Interactive image segmentation via backpropagating refinement scheme , author=. CVPR , pages=

  24. [32]

    CVPR , pages=

    F-BRS , author=. CVPR , pages=

  25. [33]

    arXiv:2102.06583 , year=

    Reviving iterative training with mask guidance for interactive segmentation , author=. arXiv:2102.06583 , year=

    arXiv

  26. [34]

    ICCV , @pages=

    Conditional diffusion for interactive segmentation , author=. ICCV , @pages=

  27. [35]

    ICPR , @pages=

    EISeg: Effective interactive segmentation , author=. ICPR , @pages=. 2016 , organization=

    2016

  28. [36]

    TIP , @volume=

    Neutro-connectedness cut , author=. TIP , @volume=. 2016 , publisher=

    2016

  29. [37]

    CVPR , @pages=

    Interactive object segmentation with inside-outside guidance , author=. CVPR , @pages=

  30. [38]

    arXiv:1707.00243 , year=

    Deep GrabCut , author=. arXiv:1707.00243 , year=

    arXiv

  31. [39]

    IEEE JSSC , volume=

    Design of an image edge detection filter using the Sobel operator , author=. IEEE JSSC , volume=. 1988 , @publisher=

    1988

  32. [40]

    DLMIA , pages=

    Generalised Dice Overlap , author=. DLMIA , pages=. 2017 , organization=

    2017

  33. [41]

    CVPR , pages=

    Boundary IoU , author=. CVPR , pages=

  34. [42]

    MedIA , @volume=

    CHAOS challenge-combined (CT-MR) healthy abdominal organ segmentation , author=. MedIA , @volume=. 2021 , @publisher=

    2021

  35. [43]

    CVPR , @pages=

    Interactive segmentation as gaussion process classification , author=. CVPR , @pages=

  36. [44]

    CVPR , @pages=

    Focused and collaborative feedback integration for interactive image segmentation , author=. CVPR , @pages=

  37. [45]

    CVPR , @pages=

    Efficient mask correction for click-based interactive image segmentation , author=. CVPR , @pages=

  38. [46]

    ACMMM , pages=

    Feature decoupling-recycling network for fast interactive segmentation , author=. ACMMM , pages=

  39. [47]

    MedIA , @volume=

    Automated Segmentation of knee bone and cartilage combining statistical shape and CNN , author=. MedIA , @volume=. 2019 , @publisher=

    2019

  40. [48]

    ECCV , @pages=

    Learning to predict crisp boundaries , author=. ECCV , @pages=

  41. [49]

    ICCV , @pages=

    Adaptis: Adaptive instance selection network , author=. ICCV , @pages=

  42. [50]

    CVPR , @pages=

    Feature pyramid networks for object detection , author=. CVPR , @pages=

  43. [51]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    clDice-a novel topology-preserving loss function for tubular structure segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.