Pith. sign in

REVIEW 2 major objections 6 minor 13 references

Weak support masks contaminate few-shot segmentation; cleaning them with sparse-atom evidence recovers reliable class regions before any query is seen.

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 · grok-4.5

2026-07-31 07:14 UTC pith:VC5U6IBC

load-bearing objection Solid plug-in cleaner for weak-support FSS; trust the Raw comparisons and ablations more than the headline SAM3 gaps. the 2 major comments →

arxiv 2607.24706 v2 pith:VC5U6IBC submitted 2026-07-27 cs.CV

SADe: Sparse-Atom Support Decontamination for Few-Shot Segmentation with Weak Support Annotations

Hang Xing , Guangjun Liu , Yan Xia , Xueming Ding This is my paper
classification cs.CV
keywords few-shot segmentationweak support annotationssupport decontaminationsparse autoencoderatom evidencereliability routerprototype cleaningpredictor-agnostic
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.

Few-shot segmentation usually assumes clean pixel masks on the support images, but real annotations are often boxes, scribbles, or coarse masks that also grab background and look-alike texture. That junk gets baked into prototypes and prompts and hurts the query. SADe is a small, support-only cleaner: it scores which patches inside the weak mask are trustworthy, then hands a cleaned mask to any downstream model without touching query inference. Its distinctive signal is sparse-autoencoder atom evidence—contrasting which sparse factors fire inside versus outside the weak region—combined with ordinary dense similarity and a light router trained once on synthetic weak episodes from FSS-1000 and then frozen. Under matched protocols the cleaned masks beat raw weak support on most box-family settings, stay competitive with SAM3-derived masks on tight boxes, and pull far ahead when boxes are expanded. A sympathetic reader cares because the method turns messy practical annotations into reusable support evidence without retraining each segmenter.

Core claim

SADe shows that support-side reliability can be estimated without query features by fusing dense similarity with sparse-autoencoder atom contrast (inside versus outside the weak support), then routing those signals into a cleaned binary mask. That mask, produced once per support image by a frozen router trained only on synthetic FSS-1000 weak episodes, improves query mIoU over raw weak support in the large majority of matched box-family plug-in comparisons across frozen heterogeneous predictors and transfers across datasets.

What carries the argument

SAE atom evidence: for each sparse dictionary atom, contrast mean activation in the weak foreground versus its complement, form a signed contrast weight, project patch activations onto the resulting evidence vector, and calibrate against the complement distribution—then fuse with dense prototype margin scores in a lightweight reliability router.

Load-bearing premise

The method assumes that contrasting sparse factors inside versus outside a weak support region, with a router trained only on synthetic weak masks from one dataset, yields transferable reliability labels on real weak supports in other datasets and domains without ever seeing the query.

What would settle it

Run the same frozen router and SAE pipeline on held-out box-r2/box-r4 episodes where texture-similar distractors dominate the weak box; if cleaned masks no longer raise query mIoU over raw support across the four frozen predictors, or if zeroing the top target-evidence atoms no longer drops performance, the central claim fails.

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

If this is right

  • Weak boxes, loose boxes, and coarse masks can be turned into standard cleaned support masks and reused by heterogeneous frozen FSS models without changing their query paths.
  • Gains grow as the weak region admits more context, so expanded-box and coarse-annotation settings benefit most.
  • A single router trained on synthetic FSS-1000 weak episodes can be frozen and applied across COCO, PASCAL, LVIS, aerial, and medical settings when dense and atom evidence are rebuilt from unlabeled features.
  • Atom evidence supplies non-redundant reliability signal: fixed fusion and learned routing both beat dense-only, and removing top target-evidence atoms hurts query mIoU.

Where Pith is reading between the lines

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

  • Any pipeline that builds class prototypes or visual prompts from noisy spatial supervision could insert the same support-only cleaner before aggregation.
  • Because the router consumes scalar atom scores rather than atom identities, independently trained dictionaries on new domains need no alignment step—suggesting cheap domain swaps for other sparse-factor tools.
  • Point and scribble prompts barely move, implying the method’s value is mainly decontamination rather than mask completion when little context is present.

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 / 6 minor

Summary. The paper formulates weak-support few-shot segmentation as a support-annotation cleaning problem and introduces SADe, a predictor-agnostic module that scores support-patch reliability without query access. Its key signal is SAE atom evidence: a sparse autoencoder over frozen DINOv3 patch features is used to contrast atom activations inside the weak foreground vs its complement, yielding factor-level cues intended to separate target from texture-similar context. A small router (44.8k parameters) fuses atom evidence, dense prototype similarity, and episode statistics, and projects the result to a cleaned binary mask. The router is trained once on synthetic weak masks derived from FSS-1000 and frozen for all target evaluations; the SAE and PCA are fitted on unlabeled COCO features. Empirically, the paper reports: (i) standalone results under a shared ProMi query head, where SADe wins 6 of 9 prompt-shot cells on COCO-20i; (ii) paired Raw/SADe plug-in gains in 70 of 72 box-family cells across four frozen predictors on COCO-20i and PASCAL-5i; (iii) headline margins of +11.17/+19.49 mIoU over SAM3-guided ProMi under box-r2/box-r4; and (iv) appendix extensions to LVIS-92i, iSAID-5i, and a Kvasir-SEG to CVC-ClinicDB medical task, plus ablations, a lambda_res sensitivity sweep, and an atom-removal control.

Significance. If the results hold, this is a useful and well-engineered contribution: the decoupling of support cleaning from query inference is conceptually clean and practically attractive, and the SAE foreground/complement atom contrast is, to my knowledge, a new signal for patch reliability. The empirical package is a genuine strength worth naming: matched standalone controls sharing one frozen ProMi head, paired Raw/SADe plug-in cells across four heterogeneous frozen predictors, multi-shot and multi-prompt grids, cross-domain transfer to aerial and medical imagery with the router frozen, component ablations, a lambda_res sensitivity sweep, a target-evidence atom-removal control, and support-side latency/memory measurements. The central claim (support-only cleaning helps under box-family prompts, transfers frozen) is supported independently of any single comparator, which makes it robust. Significance is moderated by the fact that gains concentrate in the expanded-box regime where cleaning has something to remove, and by unresolved questions about the SAM3 comparator used for the headline numbers.

major comments (2)
  1. [§Standalone Weak-Box FSS, Table 1; Tables F.1-F.6] The abstract's signature numbers (11.17/19.49 over SAM3-guided ProMi) rest on a baseline that is anomalously below the do-nothing control under loose boxes: on COCO-20i box-r4, SAM3-guided ProMi averages 14.50 vs raw-mask ProMi's 25.10, and the same inversion holds on PASCAL-5i (40.20 vs 47.01, Table F.1), LVIS-92i (10.70 vs 17.02, Table F.2), and iSAID-5i (23.33 vs 31.59, Table F.3). A box-prompted segmenter given a box fully containing the target should rarely produce a support input worse than the expanded box itself. The described interface (SAM3 pixel mask -> box restriction -> rasterization onto the 32x32 token grid with a coverage threshold) has several places where systematic erosion can enter, and no support-side mask quality is ever reported for the SAM3 branch, so the widening gap is asserted, not shown. Please report support-mask IoU vs GT for the Raw / SAM3-guided / SADe con
  2. [Tables 1, 2, F.1-F.11; Abstract] The cleaning is not uniformly safe under low-contamination prompts, and the abstract understates this. Under tight boxes, standalone SADe is below raw-mask ProMi in all six dataset configurations, by up to 8.7 points (PASCAL-5i box Avg 55.35 vs 64.07, Table F.1; medical box Avg 39.11 vs 46.23, Table F.5). On sparse prompts, INSID3/PASCAL-5i loses at all three shots (e.g., 10-shot 41.37 vs 47.80, Table 2), so 'remains close to the corresponding raw-support baseline' is an average-level claim that hides per-cell regressions of 2-6 points. The likely mechanism -- over-pruning when the weak support is already pure, plausibly interacting with the 0.5 floor and Q0.35 cutoff in Eq. (15) -- is never analyzed. Please add a failure-mode discussion (e.g., episode-purity vs delta analysis), report point and scribble separately, and qualify the abstract/intro wording to state the operating regime in
minor comments (6)
  1. [§Experiments] All comparisons are fully paired by construction, yet no uncertainty estimates are given anywhere. Paired bootstrap confidence intervals on the Raw/SADe deltas in Table 2 (and the standalone matched tables) would be cheap and would strengthen the 70/72 and 205/216 cell counts.
  2. [Method; Table B.1] Sensitivity is shown only for the residual coefficient lambda_res (App. D). Other fixed constants directly determine the mask or the labels -- the 25% clean-foreground coverage threshold for patch labels, the 0.5/Q0.35/Q0.95 constants in Eqs. (14)-(15), Ka=128, the 0.25/0.75 PCA fusion weights, the 80% atom-exclusion frequency, and the mixing-target schedule (c_alpha=0.52, beta_alpha=8). Brief sensitivity or justification for at least the projection constants and the coverage threshold would help, since they mediate the tight-box regressions noted above.
  3. [Table 2] In Table 2, 'sparse' averages point and scribble prompts, which have quite different contamination profiles. Please report the two separately, at least in the appendix.
  4. [Tables F.10-F.11; §F.2] In the medical plug-in tables, several FSS-SAM3 cells sit near zero (e.g., Table F.10: 0.00->3.67, 2.85->1.05). These cells count equally in the 205/216 improved-cell tally and the +8.66 mean gain. Consider excluding cells with near-zero Raw performance from the headline counts, or noting their influence.
  5. [Figure 2; §Introduction] The Introduction describes Figure 2 as comparing 'Dense and SADe score maps', but the caption presents violin plots of dense prototype margin z-scores before/after reweighting. Please align text and caption; also note the caption of Figure 1 would benefit from stating which column shows atom vs dense evidence.
  6. [Eq. (8); Table F.2] Eq. (8): when the weak-complement score distribution is nearly degenerate, Q0.90 - Q0.10 can approach zero; please state the guard behavior. Also, one line explaining why LVIS-92i uses the FSS-1000 unlabeled source rather than COCO (Table F.2) would preempt confusion.

Circularity Check

0 steps flagged

No significant circularity: support cleaning is trained on synthetic FSS-1000 labels and evaluated by external query mIoU on held-out datasets with a frozen router.

full rationale

SADe’s load-bearing chain is ordinary supervised transfer, not a closed definitional loop. Patch reliability labels ys_i are derived from clean-foreground coverage on synthetic weak masks from FSS-1000; the router and SAE/PCA are then frozen. Reported results are query mIoU against independent query GT on COCO-20i, PASCAL-5i, and other held-out sets, under matched protocols that do not feed query features into cleaning. Dense and atom scores are constructed from support-side features only; atom evidence is a foreground/complement activation contrast, not a quantity defined by the downstream mIoU. Ablations (Dense-only vs Atom-only vs full; target-evidence atom removal) and plug-in reuse across frozen predictors further test complementary content rather than rename a fitted input as a prediction. No uniqueness theorem, self-citation chain, or ansatz-via-citation forces the central claim. Skeptical concerns about the SAM3-guided ProMi baseline under expanded boxes are experimental-comparator issues, not circularity of the derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on standard CV/ML scaffolding (frozen DINOv3 features, SAE dictionary learning, prototype cosine margins) plus paper-specific constructions for atom reliability and a heavily parameterized router/projection stack trained on synthetic weak supports. No new physical entities; the invented pieces are methodological objects whose only external handle is downstream mIoU.

free parameters (7)
  • Router mixing-target schedule (cα=0.52, βα=8, α∈[0.05,0.95], α0=0.35) = cα=0.52, βα=8, bounds [0.05,0.95], init 0.35
    Hand-fixed mapping from weak-support purity ρ to episode mixing target α*; directly shapes how much the learned router overrides dense confidence.
  • Patch label FG coverage threshold (25%) = ≥25% clean-foreground coverage
    Defines ys_i supervision inside Sw during router training; changes which weak patches count as reliable targets.
  • Reliability projection quantiles and floors (Q0.95 norm, max(0.5, Q0.35), retry rules) = 0.95 / 0.35 / 0.5 floor with standalone/plug-in retries
    Episode-adaptive binarization of ri into the cleaned mask; controls precision/recall of retained support tokens.
  • Episode-salient atom count Ka and SAE TopK budgets = Ka=128; 16k/32 eval; 8k/24 train
    Selects which atoms enter u and how sparse codes are formed; fixed rather than derived (Ka=128; eval SAE 16k/Budget32; train SAE 8k/Budget24).
  • Dense feature fusion weights and PCA rank (0.25 xi + 0.75 P128) = rank-128, mix 0.25/0.75
    Stabilizes dense prototypes; rank and mix coefficients are design choices affecting sd_i.
  • Loss weights λs=1.0, λα=0.25 and router train budget (800 steps) = λs=1.0, λα=0.25, 800 AdamW steps
    Balance patch BCE vs episode mixing BCE and early-stop/selection regime on FSS-1000.
  • High-frequency atom exclusion threshold (80% activation) = exclude atoms active in ≥80% of unlabeled patches
    Drops near-ubiquitous atoms before evidence construction; changes the atom dictionary effectively used at test time.
axioms (5)
  • domain assumption Frozen self-supervised patch features (DINOv3) form a class-agnostic space in which support reliability can be estimated without task-specific finetuning.
    Stated in Introduction/Related Work and used throughout feature extraction; standard FSS foundation-model assumption.
  • ad hoc to paper Contrasting mean SAE atom activations in weak foreground vs complement yields factor-level target-vs-context cues not captured by dense similarity to a noisy reference.
    Core methodological hypothesis in Sparse Atom Bank; supported by ablations but not independently proven.
  • ad hoc to paper A router trained on synthetic weak masks from FSS-1000 transfers, frozen, to COCO/PASCAL/LVIS/aerial/medical supports and multiple predictors.
    Explicit train-once/freeze-all protocol; load-bearing for predictor-agnostic and cross-domain claims.
  • domain assumption Support cleaning can be fully decoupled from query inference: a binary cleaned mask through native interfaces is a sufficient interface to heterogeneous FSS models.
    Interface definition (Eq. 1) and plug-in protocol; assumes downstream models benefit from better masks without joint training.
  • domain assumption BatchTopK sparse autoencoders provide a stable, reusable dictionary over unlabeled DINOv3 features for evidence construction.
    Uses prior SAE methodology (Makhzani & Frey; Bussmann et al.) as infrastructure.
invented entities (2)
  • SAE atom evidence score ai (signed foreground/complement atom contrast projected per patch and calibrated on Bw) no independent evidence
    purpose: Supply factor-level patch reliability cues complementary to dense prototype margins.
    Defined in Eqs. 4–8; central new evidence channel. Falsifiable only via ablations/downstream mIoU, not via an external predicted observable.
  • Reliability router Rθ / α(E) producing cleaned mask Ms̃ no independent evidence
    purpose: Fuse atom, dense, membership, and episode statistics into a predictor-agnostic support mask.
    Lightweight MLP trained on synthetic episodes; the operational object shipped as SADe.

pith-pipeline@v1.2.0-grok45-kimik3 · 28752 in / 4099 out tokens · 72602 ms · 2026-07-31T07:14:15.357624+00:00 · methodology

0 comments
read the original abstract

Few-shot segmentation (FSS) commonly assumes clean pixel-level support masks, yet practical support supervision often uses boxes, scribbles, coarse masks, or pseudo-masks. These weak annotations may include texture-similar distractors and background context alongside the target, contaminating class prototypes or visual prompts before query prediction. We introduce SADe, a predictor-agnostic support decontamination layer that estimates the reliability of selected support patches without query information. Central to SADe is sparse autoencoder (SAE) atom evidence: dense similarity may respond to both target and texture-similar context, whereas contrasting atom activations inside and outside the weak-support region provides factor-level reliability cues. A lightweight router combines atom evidence with dense similarity and episode statistics to predict patch reliability and generate a cleaned support mask. Trained once on synthetic weak-support episodes from FSS-1000, the router is frozen for all target evaluations. The resulting mask supports standalone prediction or can be supplied to heterogeneous FSS models through native support interfaces without altering query-side inference. Under a matched weak-support protocol, SADe achieves the highest query mIoU in six of nine standalone prompt-shot combinations. With the same ProMi query head, it is within 0.03 mIoU of SAM3-derived masks under tight boxes and surpasses them by 11.17 and 19.49 points under box-r2 and box-r4, respectively. As a plug-in, SADe improves over raw support in 70 of 72 matched box-family comparisons across four frozen downstream models and two datasets. On point and scribble prompts, its average performance remains close to the corresponding raw-support baseline. Ablations and atom-removal controls show that atom evidence contributes reliability information beyond dense similarity.

Figures

Figures reproduced from arXiv: 2607.24706 by Guangjun Liu, Hang Xing, Xueming Ding, Yan Xia.

Figure 1
Figure 1. Figure 1: Support-side evidence and cleaning outcomes. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Support-side score distributions. Violin plots show [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of SADe. A frozen DINOv3 backbone encodes the support image into a support-token grid, to which the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: COCO-20i plug-in examples under box-r4 weak support, grouped by downstream model. Support/query GT are green, weak boxes and predictions red, and cleaned support blue. Each Raw/SADe pair shares the episode, frozen model, and query inference; only support input differs. further analyzes the Dense+Atom fusion coefficient and shows that, on both FSS-1000 and COCO-20i , this bene￾fit persists across a broad ra… 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

13 extracted references · 3 linked inside Pith

  1. [4]

    InInternational Conference on Learning Representations

    Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching. InInternational Conference on Learning Representations. Makhzani,A.;andFrey,B.J.2014.k-SparseAutoencoders. InInternational Conference on Learning Representations. Min, J.; Kang, D.; and Cho, M

    2014

  2. [6]

    InInternationalConference on Learning Representations

    SAM 2: Segment AnythinginImagesandVideos. InInternationalConference on Learning Representations. Oral. Shaban,A.;Bansal,S.;Liu,Z.;Essa,I.;andBoots,B.2017. One-Shot Learning for Semantic Segmentation. InPro- ceedingsoftheBritishMachineVisionConference(BMVC), 167.1–167.13. BMVA Press. Siméoni, O.; Vo, H. V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C....

    2017

  3. [7]

    Sun,Y.;Chen,J.;Zhang,S.;Zhang,X.;Chen,Q.;Zhang,G.; Ding, E.; Wang, J.; and Li, Z

    Interpretable and Testable Vision Features via Sparse Au- toencoders.arXiv preprint arXiv:2502.06755. Sun,Y.;Chen,J.;Zhang,S.;Zhang,X.;Chen,Q.;Zhang,G.; Ding, E.; Wang, J.; and Li, Z

    arXiv

  4. [8]

    Wang, K.;Liew, J

    Few- Shot Semantic Segmentation Meets SAM3.arXiv preprint arXiv:2604.05433. Wang, K.;Liew, J. H.; Zou,Y.; Zhou, D.;and Feng, J.2019. PANet:Few-ShotImageSemanticSegmentationwithProto- type Alignment. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 9197–9206. Yao, X.; Cao, Q.; Feng, X.; Cheng, G.; and Han, J

    Pith/arXiv arXiv 2019

  5. [9]

    InPro- ceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, 73918–73956

    Interpret- ing CLIP with Hierarchical Sparse Autoencoders. InPro- ceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, 73918–73956. PMLR. Zakir,H.M.;andHo,E.T.W.2026. RevealingtheSemantic Selection Gap in DINOv3 through Training-Free Few-Shot Segmentation.arXiv preprint arXiv:2602.07550....

    arXiv 2026

  6. [10]

    standalone comparison and paired plug-in protocol of the main paper. Building on this protocol,weaddaPASCAL-5 istandalonecomparisonunder the shared ProMi query head and evaluate broader natural- image categories on LVIS-92i, aerial remote-sensing im- agery on iSAID-5i, and cross-dataset medical endoscopy with Kvasir-SEG support and CVC-ClinicDB queries. C...

    2014

  7. [12]

    budget of 32; the FSS-1000 source uses the 8k-atom, budget-24 SAE that supplies atom evidence during router training. Downstream predictor Preprocessing FSSDINO and SANSA Images and masks are resized directly to512×512without padding INSID3 Support masks are mapped to original-image coordinates; inference uses the native 1024-pixel resolution FSS-SAM3 Ori...

    arXiv

  8. [13]

    in all six configura- tions. Relative to raw-mask ProMi, the gains are+2.07and +2.70mIoUpointsonPASCAL-5 iandLVIS-92i,+1.51and +1.86points on iSAID-5i with COCO and iSAID sources, and+0.50and+1.02points in medical endoscopy with COCO and Kvasir-SEG sources. Under the shared ProMi query-prediction path, SADe outperforms raw-mask ProMi in all six natural-im...

    arXiv 1918

  9. [2020]

    and query images from CVC-ClinicDB (Bernal et al. 2015). Unlabeled image sources.The source labels in the re- sult tables refer to COCO train2014, FSS-1000 images, the iSAID (Zamir et al

    2015

  10. [2021]

    InProceedings of the IEEE/CVF International Conference on Computer Vision, 6941–6952

    Hypercorrelation Squeeze for Few-Shot Segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, 6941–6952. Nguyen,K.;andTodorovic,S.2019. FeatureWeightingand Boosting for Few-Shot Segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, 622–631. Oh, A.; Shin, W.; and Kim, S

    2019

  11. [2024]

    Carion, N.; Gustafson, L.; Hu, Y.-T.; Debnath, S.; Hu, R.; Suris, D.; Ryali, C.; Alwala, K

    BatchTopK Sparse Autoencoders.arXiv preprint arXiv:2412.06410. Carion, N.; Gustafson, L.; Hu, Y.-T.; Debnath, S.; Hu, R.; Suris, D.; Ryali, C.; Alwala, K. V.; Khedr, H.; Huang, A.; Lei, J.; Ma, T.; Guo, B.; Kalla, A.; Marks, M.; Greer, J.; Wang, M.; Sun, P.; Rädle, R.; Afouras, T.; Mavroudi, E.; Xu, K.; Wu, T.-H.; Zhou, Y.; Momeni, L.; Hazra, R.; Ding, S....

    Pith/arXiv arXiv 2025

  12. [2025]

    InEuropeanConferenceonArtificialIntel- ligence, 4016–4023

    Label Any- thing: Multi-Class Few-Shot Semantic Segmentation with VisualPrompts. InEuropeanConferenceonArtificialIntel- ligence, 4016–4023. IOS Press. Gupta,A.;Dollár,P.;andGirshick,R.2019. LVIS:ADataset forLargeVocabularyInstanceSegmentation.InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 5356–5364. Han,B.;andOh,T.-...

    2019

  13. [2026]

    Jha, D.; Smedsrud, P

    RUFNet: Query-Guided Support Mask Refinement and Uncertainty Fusion Based on Hybrid Mamba for Few-Shot Brain Tumor Segmentation.arXiv preprint arXiv:2607.05035. Jha, D.; Smedsrud, P. H.; Riegler, M. A.; Halvorsen, P.; deLange,T.;Johansen,D.;andJohansen,H.D.2020.Kvasir- SEG:ASegmentedPolypDataset. InMultiMediaModeling, 451–462. Kirillov, A.; Mintun, E.; Ra...

    Pith/arXiv arXiv 2020