acceptodds
Withdrawn from ICLR 2027

Controlling Spurious Foreground Expansion in Open-Vocabulary Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Open-vocabulary segmentation identifies image regions corresponding to text queries, even when the queried classes are absent from the image. During single-image test-time adaptation, confidence sharpening can suppress true foreground, while anchor self-training can turn a weak false foreground into its own supervision. The resulting update enlarges the false mask, which then supplies more anchors. This second failure is termed spurious foreground expansion (SFE). To control this feedback loop, the proposed Spurious Foreground Expansion Guard (SFE-Guard) provides a two-stage guard. Credit-Denominated Issuance (CDI) freezes zero-shot semantic evidence and bounds the pseudo supervision issued to each foreground class. Hierarchical Risk Correction (HRC) reviews aggregate and background risks and applies a feasible correction before a candidate update proposed by the optimizer is committed and reused as supervision. Across all three datasets, SFE-Guard improves state-of-the-art Present Dice by 3.066 and 3.177 points for one and multiple queries, respectively. It removes 77.9% of the excess false foreground caused by anchor-based self-training in one-query OVS and restores multi-query absent class area to the zero-shot level.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Open-vocabulary segmentation, test-time adaptation, spurious foreground expansion, pseudo-label self-training, dynamic anchors, risk-constrained optimization.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_seTiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SFE-Guard, which combines pseudo-label weighting and selection based on frozen zero-shot evidence with constrained update correction to limit false foreground expansion during test-time adaptation for open-vocabulary segmentation.

Critical strengths
  • The paper reports improvements under its evaluation protocol. The component ablations in Table 2 appear controlled and provide evidence that CDI and HRC contribute to the reported Present Dice gains.
  • The paper addresses a useful problem: test-time self-training can amplify false foreground predictions when queried classes are absent. SFE-Guard combines frozen zero-shot evidence for weighting and selecting pseudo-labels (CDI) with constrained correction of proposed updates (HRC).
Critical weaknesses
  1. The primary metric does not directly penalize a central failure the paper aims to address. “Present Dice” averages Dice only over images in which a class is present. Consequently, false predictions of that class on absent images do not reduce the headline score. For example, perfect segmentation on one present image and an equally large false mask on an absent image would yield Present Dice of 100%, while dataset-level class IoU would be 50%. Reporting absent foreground area separately is informative, but does not establish superior overall semantic segmentation performance. Standard mIoU results are needed to determine whether the reported improvements hold under established evaluation.
  2. The spatial evaluation protocol departs materially from the cited methods. The submission resizes entire images to 224×224. In contrast, NACLIP uses a shorter side of 336, or 560 for Cityscapes, followed by sliding-window inference; MLMP reconstructs full-resolution Cityscapes predictions from overlapping patches. Compressing an entire street scene into a small square changes object scale and aspect ratio, potentially affecting small objects and boundaries. Although applying the same preprocessing across methods supports internal comparisons, it does not establish comparability with published segmentation benchmarks. The authors should justify this departure and evaluate the method under an established protocol.
  3. The task and benchmark coverage are insufficient to support the broad open-vocabulary segmentation claims. One-query binary segmentation is a reasonable diagnostic for absent-query behavior, but it merges all nonqueried categories into background and removes much of the semantic competition required for scene parsing. The paper also includes joint multi-query evaluation, but its COCO setting covers only 80 object categories. It does not evaluate full COCO-Stuff with 171 categories or ADE20K with 150 categories, despite these benchmarks being used by the underlying NACLIP framework. Since the formulation supports arbitrary class counts, the omission requires justification. Joint evaluation on these benchmarks with standard mIoU would directly test whether the method generalizes beyond the reported diagnostic settings.
  4. The choice of BAPA as the principal methodological foundation needs stronger justification. The concern is the narrow scope of its binary adaptation formulation. The paper should clarify why this formulation provides an appropriate foundation for broader open-vocabulary semantic segmentation, and distinguish category-conditioned binary segmentation from referring expression segmentation. It should also establish whether the observed improvements require the proposed mechanisms or can be obtained through simpler supervision weighting and conservative updates.
  5. The following issues should be corrected, although they are secondary to the evaluation concerns above:
  • Appendix E expands CMS as “counterfactual mask score.” The metric introduced in the cited Counterfactual Segmentation Reasoning paper is Confusion Mask Score.
  • Tables C.2–C.3 introduce “Senior background” and “Account,” whereas the method defines background and aggregate risks without explaining this additional terminology.
  • “Credit-Denominated Issuance” uses a banking metaphor for confidence-dependent weighting and anchor selection. A more direct description would make the methodological contribution easier to assess.
  • Page 24 ends with the isolated sentence, “prediction artifacts will accompany the submission,” which appears to be an unfinished editorial fragment.
Questions for authors

I think this paper requires major revisions. With solid experiments on widely used benchmarks and settings provided, I would consider changing my ratings.

3 comments

Thank you for the evaluation suggestions. We address standard metrics, spatial protocols, and full scene-parsing datasets below. Scores are percentages.

W1–W3 / Questions for Authors: Standard evaluation.

Standard mIoU. The additional results below include standard mIoU, computed from the dataset-level confusion matrix. The NACLIP table also reports Present Dice, absent area, and pixel accuracy.

Spatial protocol and comparison with published OVS models. Beyond the original 224x224 evaluation, we added aspect-preserving resizing and original-resolution scoring. On the full VOC validation set, SCLIP retains its official 80-template ensemble, synonyms, sliding-window inference, temperature, and postprocessing. Adaptation uses one class residual shared across synonyms and normalized class scores; zero-shot predictions exactly reproduce the official inference path. Official SCLIP zero-shot obtains mIoU/Present Dice of 59.691/62.873. Multi-BAPA obtains 27.706/49.780, CDI-only 48.333/61.170, and CDI+HRC 62.384/64.241. Compared with zero-shot, CDI+HRC improves mIoU and Present Dice by 2.693 and 1.368 points, respectively.

Original-resolution NACLIP evaluation. On all 1,449 VOC validation images, NACLIP ViT-L/14 uses aspect-preserving resizing, original-resolution scoring, our three templates and last-12-layer mean features.

NACLIP on the full VOC validation set

MethodmIoUPresent DiceAbsent areaPixel accuracy
Zero-shot29.70751.6872.30532.347
Multi-BAPA7.03940.0944.49014.504
CDI-only24.26152.5702.92750.063
CDI+HRC31.52458.6691.95956.752

CDI+HRC improves all four metrics over Multi-BAPA, CDI-only, and zero-shot. Compared with zero-shot, mIoU and Present Dice increase by 1.817 and 6.982 points, respectively, while absent area decreases from 2.305% to 1.959% and pixel accuracy increases from 32.347% to 56.752%. The full model improves segmentation quality while reducing false foreground.

ADE20K and COCO-Stuff. We evaluated the full validation sets: 2,000 images/150 classes for ADE20K and 5,000 images/171 classes for COCO-Stuff. The table reports mIoU.

DatasetZero-shot mIoUMulti-BAPACDI-onlyCDI+HRC
ADE20K16.3426.59213.96618.243
COCO-Stuff20.5195.68518.74321.233

CDI+HRC achieves the highest mIoU among the compared methods on both datasets, exceeding Multi-BAPA by 11.651/15.548 points and zero-shot by 1.901/0.714 points. Adding HRC improves CDI-only by 4.277/2.490 points, demonstrating complementary benefits of supervision control and update correction. These results show that the full model improves self-training and outperforms the frozen model in scene parsing with larger class sets.

W4: Choice of anchor self-training and component roles. We chose BAPA for its observation that severe class imbalance can drive existing adaptation methods to collapse, particularly in binary segmentation. We consider its class-balanced anchor strategy a sound response to this problem. Although multi-query segmentation has broader applications, we view the binary setting under severe class imbalance as a useful stress test of whether adaptation can preserve a small foreground against a dominant background. We adopt anchor self-training because class-balanced anchor supervision alleviates foreground suppression. Its dynamic refresh also exposes the problem we address: erroneous predictions become later supervision and drive false foreground expansion. CDI limits each class's total supervision using fixed semantic evidence, while HRC corrects candidate updates to reduce the risk of reinforcing prediction errors.

On full VOC, we compared simpler confidence weighting and its combination with KL consistency. Confidence weighting obtains Present Dice of 43.048%; combining it with KL consistency gives a best result of 56.233% across coefficients 0.1, 1, and 10, versus 58.669% for CDI+HRC. This gain over simpler alternatives demonstrates the value of class-total supervision control and candidate-update correction.

W5: We apologize for overlooking this stray sentence during the appendix checks. AI assistance in this editing step was limited to wording; the authors rigorously reviewed the contributions and results. The sentence does not affect our core claims and will be removed in the revision. We will expand CMS as Confusion Mask Score, replace "Senior background"/"Account" with background/aggregate risk, and describe CDI as limiting class supervision using semantic evidence.

Thanks for providing the requested experiments. They support this paper better in showing the ablation effects. However, at this moment, I still have some major concerns:

  1. The overall presentation is extremely difficult to follow. Many figures and visualizations are not self-contained. The notation system is chaos and hard to follow without a table of notations. The motivations are poorly explained while a bunch of terms and theories are thrown at readers. The qualitative results are limited in that no SOTA works' performance is demonstrated.
  2. Concerns toward reimplementations: the provided supplementary material contains some code but does not cover the full pipeline. There are many details missing. For example, NaCLIP's official GitHub repo reports mIoU of 64.07 on VOC21. But the rebuttal shows "zero-shot 29.707". It is unclear to me what makes this difference or where "last-12-layer mean features" comes from.
  3. The contributions still remain incremental. The paper does not sufficiently establish what substantial conceptual advance or broadly useful capability this combination provides.

Thank you for acknowledging the additional experimental evidence and explaining your remaining concerns. We appreciate the opportunity to clarify these points.

1. Presentation and readability.

We apologize for the difficulty in following the manuscript. Given the number of equations, omitting a notation table was an oversight on our part. We believe the motivation is meaningful: self-training can turn prediction errors into later supervision and progressively amplify false foreground. The reported results and component analyses support the value of addressing this feedback process, but we recognize that its explanation needs to be clearer. We will improve the language and organization, add a notation table, clarify figure labels and captions, and supplement qualitative comparisons with official-configuration OVS predictions.

2. Implementation settings and official NACLIP comparison.

Last-12-layer mean features. Our implementation builds on MLMP's multilayer feature code to establish a consistent adaptation comparison setup. We average the outputs of the last 12 visual layers to obtain dense features, using ViT-L/14 and three templates per prototype. We should have distinguished this more clearly from the complete official NACLIP configuration. The original NACLIP paper reports 57.9% VOC21 mIoU with ViT-L/14, the same backbone used in our study; the repository's 64.07% corresponds to ViT-B/16.

The performance gap primarily reflects our simplified prompt configuration. We use three rather than 80 templates, replace predefined background concepts with a generic background prototype, and omit PAMR postprocessing. For example, the official background vocabulary includes sky, grass, wall, and road, whereas we average "a background scene," "a background photo," and "an image background." Our intention was to use a unified configuration across all test datasets without enumerating background categories separately for each one, since deployment backgrounds may extend beyond a predefined vocabulary. We intended this setting to reflect such practical conditions. Using shared experimental conditions provides a fair basis for comparing the methods and assessing their effectiveness within this setting.

To address your concern, we reevaluated CDI+HRC under the official settings. Scores are percentages.

MethodmIoUPresent DiceAbsent areaPixel accuracy
Official-configuration zero-shot57.85167.0980.18785.556
+ CDI+HRC60.24369.2650.18188.765

Compared with zero-shot, CDI+HRC improves mIoU by 2.392 points, Present Dice by 2.167 points, and pixel accuracy by 3.209 points, while reducing absent area from 0.187% to 0.181%. These results support the method's effectiveness under the official configuration as well as our simplified setting.

3. Conceptual contribution.

We appreciate the concern about the extent of the contribution. To our knowledge, this is the first framework to explicitly investigate and jointly address the distinct adaptation failures across both single-query and multi-query OVS: foreground suppression under confidence sharpening and spurious foreground expansion under anchor self-training. Although the method builds on existing techniques, we believe its contribution goes beyond combining components: it identifies this feedback problem across query settings and introduces coordinated controls at its two stages.

Reviewer_4MKNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes controlling Spurious Foreground Expansion Guard in test-time adaptation for open-vocabulary segmentation, which consists of Credit Denominated Issuance to constrain pseudo-supervision and Hierarchical Risk Correction to regulate model updates. Experiments demonstrate improvements under both single-query and multi-query settings.

Critical strengths
  1. The motivation is clear. This paper analyzes the failure mode of test-time adaptation.
  2. Extensive experimental analysis are implemented to evaluate the effectiveness.
Critical weaknesses
  1. The methodological novelty needs further clarification. CDI bounds pseudo-label supervision using frozen semantic evidence, while HRC corrects candidate updates through constrained optimization. These designs are closely related to confidence-weighted self-training and gradient projection. The authors should clarify the substantive differences and compare against simpler combinations of supervision weighting and prediction-consistency regularization to establish the benefits of the proposed formulation.
  2. CDI uses frozen semantic evidence to determine the supervision budget. However, when the initial semantic evidence is inaccurate or severely biased under domain shift, this mechanism may suppress valid foreground regions or preserve incorrect supervision. The paper should investigate performance under different levels of initial prediction quality.
  3. Table 2 shows that, in the multi-query setting, adding HRC improves present dice but reduces pixel accuracy and does not further reduce absent-class area. This indicates a trade-off across evaluation metrics in the benefits of HRC.
  4. The method uses a fixed configuration of 20 adaptation steps, a 20% anchor ratio, and a reference that protects half of the foreground probability increase in the single-query setting. The authors should analyze sensitivity to the number of steps, anchor ratio, and protection ratio, particularly whether spurious foreground expansion re-emerges during longer adaptation runs.
  5. How many queries are provided in the multi-query experiments? The authors should evaluate different query-set sizes, and semantically similar categories in multi-query to analyze their effects on model performance.
Questions for authors

Please refer to weakness.

2 comments

Thank you for the detailed assessment. We address each weakness and the corresponding Questions for Authors together below. Scores are percentages, and differences are percentage points.

W1: Distinction from simpler weighting and regularization. CDI constrains the total supervision assigned to a class using frozen semantic evidence, rather than weighting each currently confident pixel independently. Growing a predicted region therefore cannot increase its supervision without that fixed bound. HRC evaluates the proposed update and computes a constrained correction to its induced risk increases, rather than adding a fixed consistency penalty to the training loss. We compared confidence weighting, confidence weighting plus KL regularization, and CDI plus KL on all 1,449 VOC validation images, using matched features, prompts, and adaptation steps. Plain confidence weighting obtains Present Dice 43.048. For KL coefficients 0.1, 1, and 10, confidence weighting plus KL obtains 56.233, 54.403, and 53.673; CDI+HRC obtains 58.669. These results demonstrate the greater effectiveness of our strategy for present-object segmentation.

W2: Dependence on initial prediction quality. Among 6,945 present image-class pairs with initial credit below 0.25, equal-domain mean Dice increases from 16.14 at initialization to 18.76 with CDI+HRC, versus 12.10 with BAPA and 17.50 with CDI-only. Thus, weak initial evidence does not necessarily prevent improvement.

W3: Dice, pixel accuracy, and absent-area trade-offs. We agree that adding HRC is not uniformly beneficial across metrics. In the matched native-resolution VOC experiment, CDI-only obtains Dice/absent area/PA of 52.570/1.927/56.063, versus 58.669/2.159/50.752 for CDI+HRC. HRC improves present-object segmentation but sacrifices some false-foreground control. Conversely, with official SCLIP text and inference settings, HRC improves all four reported metrics over CDI-only; the full results and settings are provided in our response to Reviewer ZcA7, W2/Q2. Overall, the full model improves Present Dice over CDI-only in both model settings and all four metrics on SCLIP. These gains demonstrate the practical value of HRC for improving object segmentation.

W4: On full VOC, increasing adaptation from 20 to 60 steps improves CDI+HRC Dice from 58.669% to 59.289% and reduces absent area from 2.159% to 2.105%, showing that segmentation quality and false-foreground control are maintained with longer adaptation. At anchor ratios of 10%, 20%, and 40%, Dice is 58.753%, 58.669%, and 58.366%, and absent area is 2.379%, 2.159%, and 1.657%, respectively. Dice varies by less than 0.4 points, while larger anchor ratios further reduce absent area. For one-query protection coefficients of 0.25, 0.50, and 0.75, Dice is 63.030%, 63.016%, and 62.992%, varying by less than 0.04 points. These results demonstrate stable segmentation performance within the tested parameter ranges.

W5: Query number and semantic similarity. Thank you for highlighting the influence of query sets on adaptation. The original multi-query protocol jointly queries all foreground classes: 20 for VOC, 80 for COCO objects, and 19 for Cityscapes. On all 1,449 VOC validation images, we progressively expanded random query sets from 2 to 5, 10, and 20 classes and constructed five-class sets with similar or distant semantics based on text-prototype similarity. Selection uses no evaluation labels. Unqueried classes map to background, and methods are compared within the same query set. Values below are BAPA/CDI+HRC (%).

Query setPresent Dice ↑Absent area ↓
Random 252.814/58.86830.908/18.428
Random 546.493/59.33816.135/9.012
Random 1043.532/59.3968.742/4.638
All 2040.094/58.6694.490/2.159
Similar 547.491/59.81816.484/9.530
Distant 553.447/56.63615.042/7.934

CDI+HRC achieves higher Present Dice and lower absent area than BAPA in all six settings, showing benefits across query counts and under competition between semantically similar classes. A component-level trade-off remains: for the distant five-class set, adding HRC reduces Dice from 58.168% to 56.636% but also reduces absent area from 8.960% to 7.934%. Overall, these results support the full model's ability to improve anchor self-training segmentation and suppress false foreground expansion across different query sets.

Thank you for the detailed responses and additional experiments. The authors have addressed most of my concerns. Although the additional comparisons demonstrate improvements over confidence weighting and KL regularization, the methodological novelty of CDI and HRC beyond existing supervision weighting and constrained update strategies could be further clarified. I therefore maintain my original rating.

Reviewer_ZcA7OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes how well CDI’s semantic credit distinguishes present from absent classes and shows that SFE-Guard reduces false foreground across credit levels, including confidently misclassified absent classes. It also reports endpoint foreground suppression, counting cases where a present object was missed from the start. However, it does not directly test whether present classes with low CDI credit (especially objects initially missed or barely detected) can recover during adaptation. A useful follow-up would stratify present classes by initial credit and measure their recovery after adaptation.

Critical strengths
  1. Clearly formulates an important failure mode.: The paper identifies self-feeding errors (SFE), in which incorrectly predicted regions become training signals and amplify errors in subsequent predictions. This is an interesting and clearly defined problem.

  2. Controls errors at two stages.: CDI limits the amount of category-specific learning signal, while HRC checks the risk of an update before applying it. Together, they address how false predictions affect both adaptation and subsequent predictions.

Critical weaknesses
  1. CDI may misallocate its learning budget when initial predictions are poor.: CDI fixes its credibility scores based on predictions made before adaptation. If a present category is barely or never detected at that point, it may receive too little learning signal to recover later.

  2. HRC does not fully prevent risk, and the qualitative evidence is limited.: The theoretical guarantee relies on a local linear approximation and specific assumptions. In the experiments, the proportion of adaptation steps with increased background risk reportedly rises after calibration, although the magnitude of the increase is reduced. The paper also provides few qualitative results, and the examples appear noisy. Comparisons with open-vocabulary segmentation methods such as SCLIP [1], NaCLIP [2], and CDAM [3] would help establish the method’s performance through quantitative and qualitative evaluations.

[1] Wang, Feng, Jieru Mei, and Alan Yuille. "Sclip: Rethinking self-attention for dense vision-language inference." ECCV, (2024).

[2] Hajimiri, Sina, Ismail Ben Ayed, and Jose Dolz. "Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation." Proceedings of the winter conference on applications of computer vision. (2025).

[3] Kang, Dong Un, Hayeon Kim, and Se Young Chun. "Class distribution-induced attention map for open-vocabulary semantic segmentations." ICLR, (2025).

Questions for authors
  1. Can SFE-Guard recover categories that receive low CDI credibility scores?: Please analyze categories that are present but receive low scores, especially those that are rarely or never detected in the initial predictions. Does SFE-Guard recover these categories in later adaptation steps?

  2. Does the method generalize across backbones or work as a plug-and-play component?: Could the authors apply SFE-Guard to other open-vocabulary segmentation methods, such as SCLIP [1], NaCLIP [2], and CDAM [3], and report whether its benefits remain consistent?

  3. How robust is segmentation performance under image degradation?: Please evaluate the method on images with synthetic degradations, such as noise and blur, and report how segmentation performance changes

2 comments

Thank you for recognizing the importance of self-reinforcing prediction errors and the complementary roles of CDI and HRC. We address your concerns below. All scores are percentages; differences are percentage points.

W1/Q1: CDI fixes the evidence bounding supervision while allowing predictions and anchors to evolve. Under the three-template, 224x224 one-query protocol, we stratified 22,784 present image-class pairs across the three datasets. For the 6,945 pairs with credit below 0.25, equal-domain mean Dice is 16.135% for zero-shot, 12.104% for BAPA, 17.504% for CDI-only, and 18.760% for CDI+HRC. The full method improves over zero-shot and BAPA by 2.625 and 6.656 points, respectively, supporting average recovery under low initial credit.

W2/Q2: HRC uses a local linear approximation to derive the update correction. Although this does not guarantee monotonic reduction of the exact nonlinear risk, the empirical results show substantial reductions in the magnitude of risk increases. Section 4.2 and Appendix Tables C.1–C.3 already evaluate this behavior over all 138,980 adaptation steps on 6,949 images across the three datasets, using NACLIP ViT-L/14, 224x224 inputs, three templates, and 20 steps per image in the multi-query setting. Relative to the uncorrected candidate, HRC reduces mean positive aggregate-risk excess by 84.0%, with excess measured from the accepted state before each step. On 99.983% of steps, corrected aggregate risk is no greater than candidate risk. Although background-risk increases become more frequent, their mean positive excess decreases by 84.1%. These measurements show that HRC reduces the magnitude of risk increases in the evaluated adaptation runs.

The paper already evaluates NACLIP ViT-L/14 in both query settings (Section 4.1; Tables 1 and 2). The appendix, Implementation Details, specifies the backbone and preprocessing. To test transfer beyond this model, we tested SCLIP ViT-B/16 on all 1,449 VOC validation images using its official 80-template ensemble, class synonyms, sliding-window inference, and postprocessing. A class residual is shared across its synonyms; class scores are normalized for adaptation, while final predictions retain official postprocessing. Initialization exactly matches official SCLIP zero-shot predictions.

MethodmIoUPresent DiceAbsent areaPixel accuracy
Zero-shot59.69162.8730.13887.568
Multi-BAPA27.70649.7802.41448.153
CDI-only48.33361.1700.76877.586
CDI+HRC62.38464.2410.23389.548

CDI+HRC improves all four metrics over both adapted baselines. Compared with zero-shot, mIoU, Present Dice, and pixel accuracy increase by 2.693, 1.368, and 1.980 points, respectively, while absent area increases from 0.138 to 0.233. These results support transfer to SCLIP with improved segmentation quality.

Q3: Robustness to noise and blur. On the complete VOC validation set, multi-query CDI+HRC obtains Dice 58.669 on clean images. With Gaussian noise standard deviations 0.05 and 0.10 on RGB values in [0,1], Dice is 58.951 and 57.798. With Gaussian blur sigma 1 and 2 at the resized resolution, it is 58.839 and 58.447. Performance remains close to the clean result within these tested severities.

Thank you for the additional experiments. The SCLIP experiment partially addresses my request. However, the response refers to existing NaCLIP results and adds only SCLIP, leaving CDAM unevaluated without explanation. Please provide the requested CDAM evaluation or explain why it could not be included.

My concern about the limited qualitative results and noisy examples also remains unaddressed. The response provides no additional qualitative examples or visual comparisons to resolve this concern.

Given these remaining concerns, I maintain my original rating.

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

Related papers

Discussion (0)

Sign in to comment.