acceptodds
Under review as a conference paper at ICLR 2027

Can Frozen Hyperspherical Features Guide the Selection of Pseudo Masks?

11 views

Abstract

Foundation segmenters such as SAM return several plausible masks for an unlabeled image, and a student trained on the wrong one inherits its errors. Choosing among them means querying a second large model or fitting a quality head to annotated masks. We show that a candidate can be judged by what it does to a frozen self-supervised backbone's features. Normalized DINOv2 patch features lie on a hypersphere, and a candidate mask splits that sphere in two. Based on this reading, we introduce SphereTrust, which scores each candidate by three properties of the split, the angular contrast between the two sides, the coverage of the foreground's appearance modes, and contact with the image frame, one for each of three common ways a mask fails, and ranks a pool in 0.55 s per image from the frozen features alone. On eight SAM and SAM3 candidate pools spanning camouflaged, salient, and dichotomous segmentation and camouflage under low light, SphereTrust exceeds the strongest evaluated external baseline on six pools by 1.7 to 9.3 percentage points in mean selected Dice. These comparisons include published selection rules and explicitly labeled adaptations of DSS and UCOD-MKD. On the two prompted camouflage pools, its mean selected Dice is within 0.1 percentage points of the candidate-derived DSS adaptation, with a lower catastrophic-error rate. Which cue carries the signal depends on the candidate pool. The same sphere also supports training. The leading candidates enter as a candidate set with their scores as priors, prototypes reorder them, and a cross-fitted second round completes the labels, raising weighted F by 4.5, 2.3, and 5.5 points over fixed-label training on the three MLLM anchor pools, with students competitive with published unsupervised methods on nineteen test sets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
pseudo-label selection, unsupervised segmentation, camouflaged object detection, self-supervised vision transformers, hyperspherical features, segment anything model, self-training, salient object detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JPT1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SPHERETRUST, an unsupervised segmentation method that uses SAM/SAM3 and an MLLM(QWEN-VL) as teachers and distills their capability into a student model built on a frozen DINOv2 backbone. Although SAM can output multiple candidate masks, its ability to score these masks is poor, which is a widely acknowledged limitation. The authors’ core idea is to treat the normalized DINOv2 patch features as directions on a unit hypersphere and, using hand-crafted priors, to attribute several types of mask errors to characteristic patterns of the induced feature partition (Fig. 1). Three label-free measures (angular contrast, spherical coverage, frame contact) are combined to rank the candidates, and the leading candidates are then used to train the student through candidate-set training and a cross-fitted second round. The method is validated on candidate selection over eight SAM/SAM3 pools and on downstream camouflaged, salient, dichotomous, and low-light segmentation benchmarks.

Critical strengths
  1. Interpreting the problems of masks through the hypersphere is a novel perspective. Mapping each common failure type of a candidate mask to a property of the feature partition (Fig. 1) offers a label-free and interpretable way to rank SAM’s candidates.
  2. The paper reports the data and conclusions of each experiment in detail.
Critical weaknesses
  1. The benefit of the claimed core innovation is not sufficiently demonstrated. According to the authors’ narrative, the core innovation is to interpret mask quality through hyperspherical geometry and to select among multiple candidate masks, thereby obtaining high-quality training data (SPHERETRUST ranking). However, the paper does not sufficiently demonstrate the benefit of this aspect, and the main source of gain appears to be the teacher–student training framework rather than this selection scheme. The authors themselves state (line 460, Sec. 4.2): “the total gain does not isolate the contribution of spherical geometry.” In the ablation study (Tab. 9), removing the sphere term from both rounds (i.e., the vMF prototype term Γ; the SPHERETRUST ranking R is still used) changes the weighted F by only +0.01 on SOD and −0.21 on DIS. Moreover, under the same complete training procedure, replacing the SPHERETRUST ranking with the size control or the adapted DSS rule leaves differences of only 0.2–1.0 points (Tab. 16).
  2. The narrative around the spherical scoring is obscure and does not match the title and abstract. The narrative around this innovation is rather hard to follow; in particular, the concept of “spherical scoring” is heavily wrapped in spherical terminology, which is very confusing. Judging from the title, the paper should be about a method that scores masks with the sphere. However, in the mask scoring section (Sec. 3.1), I could find almost nothing related to the “sphere” beyond L2-normalizing the patch features and computing inner products (cosine similarities). This is inconsistent with the impression, given by the title and abstract, that the sphere plays a key role in mask selection. I hope the authors can clarify this point.
Questions for authors

(Clarification) In the mask scoring of Sec. 3.1, which component relies on spherical geometry beyond L2 normalization and cosine similarity? (Clarification) Can the authors provide evidence that isolates the contribution of the spherical selection from that of the training framework, e.g., a variant in which the scoring does not use normalized directions, evaluated on both selection quality and downstream performance?

6 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for a careful and constructive reading. Both requested experiments have been run, both are now rows of the paper, and they also led us to correct a claim of our own. One note on numbering, since the revision inserts a new Tab. 4 in the main text and every appendix table moves up by one. Tab. 9 of the submission is Tab. 10, and Tab. 16 is Tab. 17. Numbers below are those of the revision.

1. What is spherical in the selection rule. The sphere is a modelling choice, which is to read direction and discard magnitude. Patch features are normalized to unit directions, μ(S)\mu(S) is the renormalized resultant of a set of them and therefore the maximum likelihood von Mises-Fisher mean direction, angular contrast is one minus the cosine similarity between the two mean directions a candidate induces, the appearance modes are k-means centroids on unit directions projected back to the sphere, and coverage assigns patches to modes by cosine. Frame contact is not of this kind. It measures contact with the outermost pixel ring of the image frame, so it is a spatial prior rather than a statistic of the directions, and the revision now says so in Sec. 3.1 and in the abstract, which describes two properties of the split together with a spatial prior. The claim behind the other two is that direction alone carries enough signal to rank mask quality, and point 2 is the experiment that tests it directly. We accept that Sec. 3.1 did not make the construction explicit, and the revision now opens with it, including ∥f^p−f^q∥2=2−2⟨f^p,f^q⟩\lVert \hat f_p - \hat f_q \rVert^2 = 2 - 2\langle \hat f_p, \hat f_q \rangle for unit vectors, so that no stronger geometric claim is implied than the one we test. What uses the sphere distributionally rather than directionally is the vMF prototype term Γ\Gamma at training time, and the reviewer reads the training ablation, Tab. 10, correctly there. Removing Γ\Gamma from both rounds changes weighted FF by −0.65, +0.01 and −0.21 points on COD, SOD and DIS. We would only ask that Γ\Gamma and R\mathcal{R} be kept separate, since the evidence under W1 concerns R\mathcal{R}.

2. Scoring without normalized directions (Q2). Following the reviewer's suggestion, we rebuilt FF and CC in raw Euclidean geometry, with arithmetic means replacing renormalized mean directions, squared Euclidean distance replacing the cosine distance in FF, and k-means in raw space. The frame-contact term EE uses no features and is unchanged. The within-image standardization is kept so that scale is not confounded, and both arms score the same masks against the same ground truth under the same eligibility filter and the same native-resolution Dice evaluation. Removing feature normalization lowers selected Dice and Top-1 and raises the catastrophic-error rate on all eight pools.

Poolsel-DiceTop-1Dice<.2AUROC of FFAUROC of CC
SAM3 COD−3.29−3.06+2.85−7.20−0.00
MLLM COD−0.39−0.91+0.62−2.46−0.72
MLLM SOD−2.45−4.37+1.19−8.28−1.79
MLLM DIS−5.76−7.48+3.20−12.22−1.22
COD10K−3.35−7.98+2.68−3.91−4.24
DUTS-TE−10.99−19.69+8.15−12.61−3.49
DIS-VD−5.78−9.46+7.10−13.15−3.27
LL-COD−3.45−8.80+3.88−3.97−4.64

Change from removing the normalization, in points. Appendix §H.1 of the revision.

Downstream the variant trains students at .7597, .7905 and .5853 of fixed-label weighted FF against our .7526, .8144 and .6047, so the normalization is worth 2.4 points on salient objects and 1.9 on dichotomous segmentation under fixed labels, and 1.2 and 0.4 under the complete procedure, leading on every seed in both columns. On camouflage it loses 0.7, which is the same regime exception as in point 4. The variant is now a row of Tab. 17 and a section of the new Appendix §H.1, which states its construction in full.

AuthorsOpenReview ↗

Official Comment by Authors (Cont.)

3. Isolating the ranking from the training framework (W1). We agree that the rule comparison as published did not settle this, and Tab. 17 of the revision adds the two rows that do. The full-procedure column cannot separate the two stages, because candidate-set training and the cross-fitted second round are designed to recover from a poor ranking. Under fixed labels the ranking acts alone.

Ranking rule used for the labelsCODSODDIS
generator confidence.7130.7598.4847
size control.7287.8009.5842
no normalization.7597.7905.5853
adapted DSS.7624.7918.5464
SPHERETRUST.7526.8144.6047

Fixed-label student quality under five ranking rules, weighted FF, three seeds. Tab. 17 of the revision.

The gap to the generator's own mask score, which is what a practitioner faces when no external selector is used, is 4.0, 5.5 and 12.0 points. The gap to the size control is 1.4 to 2.4, and to the adapted DSS rule 2.3 on SOD and 5.8 on DIS. Across the fifteen rule and regime pairs, with each regime's mean subtracted from both variables, selected Dice and the fixed-label student's weighted FF correlate at 0.94, against 0.64 under the complete procedure. These are observed correlations on fifteen points, not a fitted model. We would add that the training stage is the third contribution of the paper rather than a confounder, so the selector is measured where it acts alone, on fixed pools in Tabs. 1 and 7 and under fixed-label training here.

4. A correction the new runs forced. On camouflage our ranking is not ahead. The adapted DSS rule trains the better fixed-label student by 1.0 point, and once the prototype fit and the pseudo labels are reseeded along with the student the five rules fall within 0.3 points in no stable order. Appendix §F.3 and the new Tab. 19 of the revision report this. Tab. 18 covers the size control and the adapted DSS rule only, and its caption now says so, since under final-student-only reseeding the unnormalized rule is ahead of us on COD on the shared seed.

Thank you for the thorough response and for candidly reporting the new experimental results, which have resolved most of my concerns. Before finalizing my evaluation, I hope the authors could provide a brief analysis explaining: (1) why removing feature normalization consistently degrades selection quality across the pools, and (2) why SPHERETRUST trails two simpler rules (no normalization and adapted DSS) on downstream COD. Even brief hypotheses or empirical observations would be valuable here, especially for readers less familiar with the dataset characteristics or domain nuances. If the authors can include this discussion in their reply and incorporate it into the main text (e.g., Experiments or Limitations) or the Appendix, I would be happy to raise my rating.

Reviewer_e91FOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPHERETRUST, a label-free pseudo-mask selection method based on frozen DINOv2 features. It evaluates candidate masks using three geometric cues: angular contrast, spherical coverage, and boundary contact. The authors further introduce a two-round training framework with candidate-set supervision, prototype-based reordering, and cross-fitted pseudo-label completion. Experiments across camouflaged, salient, dichotomous, and low-light segmentation show improvements in mask selection and downstream segmentation performance.

Critical strengths

Intuitive and interpretable mask selection. The three scoring cues address different mask failure modes and provide a practical alternative to generator confidence without training an additional quality estimator.

Extensive experimental evaluation. The authors evaluate multiple segmentation tasks, candidate pools, and selection baselines. The matched comparisons and component ablations are particularly useful for understanding the method's behavior.

Effective pseudo-label training. Candidate-set supervision and cross-fitted completion improve downstream segmentation over fixed-label training, demonstrating the value of dynamically exploiting multiple pseudo masks.

Critical weaknesses
  1. Limited independent advantage of the geometric scoring rule. Although SPHERETRUST improves mask selection over several existing methods, its advantage over the simpler area-based control is inconsistent. The size control performs better on three of four unprompted SAM pools, and the score-scale-matched downstream comparison in Appendix F.3 shows nearly identical performance, with the size control slightly outperforming SPHERETRUST on SOD and DIS. This raises questions about the incremental value of the proposed spherical coverage.
  2. Method complexity and unclear contribution of individual components. The complete framework combines geometric ranking, prototype reordering, candidate-set training, cross-fitting, and a second training round. While the overall design is coherent, several components provide limited incremental gains. For example, the second round improves COD by only 0.3 points, and alternative scoring rules achieve comparable downstream performance under the same training procedure. The additional complexity is therefore not always justified by the observed improvements.
  3. Limited backbone validation. The method relies primarily on DINOv2, with additional backbone comparisons restricted to angular contrast. The complete scoring framework is not evaluated with other recent self-supervised representations, such as DINOv3, leaving its generality across visual backbones unclear.
Questions for authors
  1. Geometric scoring. What is the main advantage of spherical coverage over the simpler area-based control, given their comparable downstream performance after score-scale matching? Under which candidate-pool conditions should the geometric formulation be preferred?
  2. Method complexity and robustness. How much of the full pipeline is necessary for the reported improvements? A performance–cost comparison with a simplified training configuration would be helpful. Given the small differences in Appendix Tables 16–17, independent end-to-end runs covering prototype fitting and pseudo-label generation would also strengthen the conclusions.
  3. Backbone generalization. Does the complete scoring method generalize to other frozen visual representations, particularly DINOv3? A fixed-candidate-pool comparison across backbones would help establish whether the proposed geometric cues are broadly applicable.
  4. Presentation and implementation details. Please improve Figure 2 by explicitly labeling the key variables, intermediate outputs, and connections between ranking and student training. Figure 3 would also benefit from larger visual examples. Additionally, a brief sensitivity analysis regarding the 1,085 annotated validation images used for configuration selection would clarify the practical annotation requirements.
4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for questions that could each be settled by an experiment. All six points below are answered in the revision. One note on numbering, since the revision inserts a new Tab. 4 in the main text and every appendix table moves up by one. Tabs. 16 and 17 of the submission are Tabs. 17 and 18 below, and the reseeded comparison the reviewer asked for is the new Tab. 19.

1. When the geometric score is preferable to the area control (W1, Q1). The reviewer is right that the size control is a strong baseline, and the new runs let us say when each of the two is preferable rather than leaving the comparison at a tie. Coverage leads on all four prompted pools, by .824/.822, .787/.763, .929/.909 and .771/.755 in selected Dice, and trails on three of the four unprompted pools. Appendix §B.1 isolates the increment directly. Among candidates of one image whose areas differ by at most five percent, coverage picks the mask with greater Dice in .851, .815 and .791 of pairs on COD10K, DUTS-TE and DIS-VD, against .481, .491 and .537 for area, and after regressing out the other terms and log area the partial Spearman correlation between coverage and true Dice is +.242 on the prompted macro. Area orders an unprompted pool well because that pool runs from whole-image masks down to fragments. It cannot separate candidates that are already the same size, which is the prompted case our pipeline operates in. We should add in fairness that the size control is a diagnostic we introduced ourselves rather than a published competitor, and that under a matched within-image score scale it is level with us on salient and dichotomous data, which Sec. 4.2 and Appendix §F.3 state.

2. How much of the pipeline is necessary (W2). We agree that a method of this shape owes the reader a cost breakdown, and we have added one. Taking a full-pool student as TT, which is 2.2 to 2.9 hours on a shared H200, round one costs 1.00 TT and already recovers 93, 37 and 62 percent of the total gain, while the second round costs a further 1.05 TT.

ConfigurationCostCODSODDIS
fixed label, selection only1.00 TT.7513.8148.6048
+ candidate-set training, round one1.00 TT.7930.8232.6384
+ second round, complete procedure2.05 TT.7961.8376.6594

Cost and contribution of each stage, weighted FF. Appendix §H.3 of the revision.

3. Independent end-to-end runs (Q2). Following the reviewer's suggestion, each of three seeds now regenerates the prototype fit, both fold students, the out-of-fold maps, the round-two candidate sets and the final student, so no stage is shared across seeds. The end-to-end standard deviation is small, at most .0027 of weighted FF for any rule and regime. On salient and dichotomous data SPHERETRUST leads all four alternatives on all three seeds, with paired margins from +0.03 to +1.48 points. On camouflage the five rules lie within 0.3 points and no ordering survives the reseeding, so we make no claim there and have corrected the paper accordingly. The full table is the new Tab. 19.

4. Backbone generalization to DINOv3 (W3, Q3). Following the reviewer's suggestion, we recomputed R\mathcal{R} on DINOv3 ViT-7B/16, with FF and CC built from the new features and the foreground and background prototypes that coverage uses rebuilt from them. Every row selects with R\mathcal{R} alone, as in Tab. 1, and the training term Γ\Gamma does not enter. A DINOv2 control at the same token count controls token-grid resolution.

FeaturesSAM3 CODMLLM CODMLLM SODMLLM DIS
DINOv2 B/14, 25×25 (paper).8241.7873.9288.7714
DINOv2 B/14, 22×22 (matched).8200.7843.9275.7664
DINOv3 7B/16, 22×22.8224.7985.9174.7619

Selected Dice, candidate pools unchanged. Appendix §H.2 of the revision.

Against that control the change is −1.0 to +1.4 points, and the score still exceeds the strongest external comparator we evaluated on three of the four pools. We used the largest released DINOv3 checkpoint so that a null result could not be attributed to a weaker model. The model sizes still differ, so we read this as evidence that the measures transfer across the two backbones tested rather than as a claim about DINOv3.

AuthorsOpenReview ↗

Official Comment by Authors (Cont.)

5. Sensitivity to the 1,085 validation images (Q4). This is a fair concern about our validation budget, and we have measured it. We re-ran the Appendix A.1 selection rule on 1,000 random subsets. At 10, 25, 50 and 75 percent of the labels it recovers the reported configuration 19, 29, 43 and 61 percent of the time. In this retrospective analysis of the tested configuration grid, subsets of roughly 109 validation images recovered the reported configuration in 19% of draws, while the empirical 95th percentile of test shortfall was zero points of weighted FF at the reported precision. That percentile is an observed quantile over the draws rather than a guarantee for every draw, and it speaks to this grid rather than to configuring the method in general. Appendix §H.4, pointed to from Sec. 4.

6. Figures 2 and 3 (Q4). We agree that both figures are harder to read than they should be. Figure 2 is redrawn in the revision. The quantities of Eqs. (1) to (6) are labelled where they are produced, including arg⁡max⁡k[πik−ℓik/T]\arg\max_k[\pi_{ik} - \ell_{ik}/T] at the selection box and the cross fitting run across image folds in round two. The panel's boundary prior is renamed to the frame contact used in the text, the feature matrix that collided with FF of Eq. (1) is gone and the panel now labels the patch directions f^p\hat f_p, and the read-out panel is enlarged. The link from the ranking to training is drawn explicitly, from the ranked pool to the retained set and labelled with the order that produces it, rather than from the single selected mask. Figure 3 now shows one regime with much larger examples, and Appendix Fig. 12 keeps all three regimes, so nothing is lost. A new figure shows qualitative picks in the two domains added for Reviewer 6G24.

Thank you for the detailed rebuttal and additional experiments. The new results address most of my concerns and strengthen the paper, although the incremental benefit of the geometric design remains somewhat limited. Overall, I appreciate the authors' efforts and will maintain my rating.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for reading our response and the revision. We are glad the clarification was helpful. Please feel free to let us know if there is anything else we can further clarify.

Reviewer_6G24OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

In order to help select segmentation candidate masks generated by the SAM model, this paper proposes SPHERETRUST, which ranks masks through three attributes of the mask (angular contrast F: the angular contrast between the two sides, spherical coverage C: the coverage of the foreground’s appearance modes, and frame contact E: contact with the image frame). Experiments demonstrate the effectiveness of the proposed method.

Critical strengths

S1: Clear presentation. I can easily understand the overall paper.

S2: Good motivation. Refine masks by the masks themselves.

S3: Extensive Experiments. Various experiments verify the effectiveness.

Critical weaknesses

W1: Paragraph allocation is unreasonable.

For example, the second paragraph of the introduction has only two sentences (3 lines), and the fifth paragraph has only 4 lines.

W2: Some of the latest methods are not cited, compared or discussed.

The author mentioned in the second paragraph of the introduction: In order to solve the problem of SAM candidate mask selection, the existing mask selectors are divided into two types: one uses the consensus between generator masks, and the other uses the second large model. However, some of the latest SAM methods try to add uncertainty into the mask [1], whether this can also solve the problem of SAM mask selection?

W3: Presentation can be improved further.

Parts of the process are not clearly expressed. For example, how the segmentation mask generated by sam divides the graphics into two partitions and the background.

W4:

The experimental data set includes difficult data domains such as Camouflage, Salient, Dichotomous, Low light, etc., but the sam3 experiment also includes some other difficult domains: such as underwater, scientific graphics. For details, see the 24 image data set segmentation experiments of sam3. Can this paper also be effective on these data domains?

[1] Segment Anything with Robust Uncertainty-Accuracy Correlation, ICML26.

Questions for authors

The questions in W2, W3 and W4

3 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the constructive suggestions, and in particular for the RUAC reference, which we had not been aware of. Two new domains have been added, and both writing points are already applied in the revision.

1. Further domains (W4). Following the reviewer's suggestion, we built prompted pools on underwater images from USOD10K and thermal imagery from VT5000. Underwater images carry a colour cast and low object-to-water contrast, and on the thermal images the two sides of a candidate differ in radiance rather than in colour. Generator, eligibility filter, selectors and students are unchanged.

SelectorUSOD10K-TE, underwaterVT5000-T test, thermal
adapted DSS rule.852 / .217 / .036.817 / .208 / .058
SPHERETRUST.849 / .239 / .031.813 / .214 / .051

Selected Dice / Top-1 / rate of Dice below .2. Tab. 4 of the revision gives all ten selectors.

We trail the adapted DSS rule by 0.3 and 0.4 points of selected Dice, and among the evaluated external comparators and ourselves we have the highest Top-1 and the lowest catastrophic-error rate on both sets. Our own size control has a higher Top-1, and the oracle a lower catastrophic rate. The pools hold 989 and 2,383 eligible images at 25.7 and 26.9 candidates each, and from random selection to the best non-oracle selector the spread is 9.8 and 8.7 points of selected Dice against a pool oracle of .892 and .863, so the ranking still has room to matter. Students trained with the complete procedure reach .805 and .613 of weighted FF against .771 and .581 for selection-only training, a gain of 3.4 and 3.2 points, and SαS_\alpha, EϕE_\phi and MAE move the same way on both sets.

2. The sets we built and did not report as comparisons (W4). Three more sets were built, and we would rather report them in Appendix §H.5 than leave them out. On RMAS-TE and ZeroWaste-f the pool oracle reaches only .525 and .345 of Dice, so the generator rather than the selector bounds every row. USIS10K-TE retains 144 images once the overlap with USOD10K is removed, which is no longer a random sample of that benchmark, and TokenCut leads on it at .647 against our .633. Domains were admitted by a size gate fixed before any pool was built, a median foreground fraction of at least five percent with at most ten percent of images below one percent, which is what excluded NDD20, EORSSD and Plittersdorf.

3. Scientific graphics (W4). We have not reached this domain. The added rebuttal experiments cover underwater and thermal imagery, and they do not establish generalization across scientific imaging domains. If the reviewer has a specific benchmark in mind we will run it and report the result in this thread.

4. Uncertainty-based selectors (W2). Thank you for bringing RUAC to our attention. Appendix J.2 of the RUAC paper averages uncertainty within each mask and compares it with mask-level IoU, so RUAC does provide a mask-level score. RUAC evaluates masks produced by its own SAM2.1 pipeline using three-click prompts, whereas our fixed pools include automatic SAM proposals, box-prompted SAM proposals, and phrase-grounded SAM3 proposals. Comparing RUAC on these pools would require adapting it, which we have not done. We have clarified this distinction in Section 2. RUAC's uncertainty estimates could be useful for choosing among candidate masks. Our generator-confidence baseline, however, uses the generator's standard predicted mask score, so it does not tell us how our method compares with RUAC.

5. Paragraph allocation (W1). We agree, and the two short introduction paragraphs are merged into their neighbours in the revision.

6. How a candidate mask partitions the image (W3). We agree this was the clarification the section needed, and Sec. 3.1 now states the partition before any equation appears. A candidate mask read on the patch grid splits that grid into the mask and its complement, and the normalized features of the two sides are two collections of directions on the unit sphere. The three measures do not read the same region. Angular contrast reads the whole grid and compares the two mean directions, coverage reads the anchor region and asks how many of its foreground appearance modes the mask side reaches, and frame contact reads the outermost pixel ring of the image frame, so the first two are statistics of those collections and the third is a spatial prior.

Thank you to the authors for the response, which has clarified my confusion. I am happy to maintain my original rating and support the acceptance of this work.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for reading our response and the revision, and for supporting the paper. We are glad the clarification was helpful. Please feel free to let us know if there is anything else we can further clarify.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.