acceptodds
Withdrawn from ICLR 2027

Calibrated Concept Denoising for Open-Vocabulary Semantic Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Frozen vision-language encoders support open-vocabulary segmentation, but their dense features often mix target evidence with correlated background and compet- ing concepts. This motivates a query-dependent scoring operation: the same chan- nel can be useful for one concept and misleading for another. We propose Cali- brated Concept Denoising (CCD), which reweights visual channels for each text query while retaining dense cosine scoring. CCD transfers concept masks from SmartCLIP, refines them with a prior-preserving residual adapter, and estimates masks directly from labeled competitor errors. A labeled set of 1,000 training images per dataset supports adaptation and branch selection; the selected branch is fixed for validation. Under the released ClearCLIP evaluator, CCD improves ViT-B/16 average mIoU by 0.48 points across six benchmarks, including a 2.20- point gain on Cityscapes. Controlled experiments across three feature families and a ViT-L/14 extension characterize where query-conditioned channel reweighting helps.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-vocabulary semantic segmentation, Vision–language models, Query-adaptive inference, Feature subspace selection, Channel gating

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c1uUOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Calibrated Concept Denoising (CCD), which reweights visual feature channels conditioned on text queries for open-vocabulary semantic segmentation. It combines transferred SmartCLIP masks, residual adaptation, and masks estimated from labeled competitor errors. Using 1,000 labeled training images per dataset, CCD improves average mIoU by 0.48 points over ClearCLIP across six benchmarks.

Critical strengths
  1. Clear motivation: Query-conditioned channel reweighting offers a complementary approach to improving dense CLIP features.

  2. Broad experiments: The paper evaluates six datasets, multiple feature families, and two backbone sizes.

  3. Useful diagnostics: Correction–regression analysis reveals when masking helps or harms predictions.

Critical weaknesses
  1. Limited overall improvement: The main gain is only 0.48 mIoU points, driven largely by Cityscapes (+2.20), while ADE20K regresses. DIH-CLIP also reports stronger results on overlapping benchmarks without target-dataset labels.

  2. Potential selection bias: Multiple branches are selected using the same 1,000 labeled training images used for adaptation or mask estimation. DirectCDD’s fold-based selection is not an independent evaluation of the complete selection procedure.

  3. Unclear novelty and reproducibility: The distinction from existing concept-conditioned masking and feature calibration needs stronger evidence. The paper also lacks complete training and feature-export code, limiting verification.

Questions for authors
  1. Does CCD remain effective with a strictly independent calibration set?

  2. Can the authors demonstrate consistent gains over stronger training-free baselines?

  3. Which components provide robust benefits across datasets rather than dataset-specific improvements?

1 comment
AuthorsOpenReview ↗

Independent calibration, gains on GLA-CLIP and reproducible evidence

Dense CLIP segmentation needs both spatially coherent features and class-sensitive comparisons. CCD adapts the channel metric at the comparison stage while keeping the visual backbone and classifier fixed. The new composition experiment establishes this complementarity on GLA-CLIP: Avg6 increases from 41.62 to 44.49 (+2.87 mIoU; paired 95% CI [+2.66, +3.07]), with improvements on five of six datasets. Under extreme annotation scarcity, CCD exceeds a text-initialized linear probe by +2.93 and +5.17 Avg6 at 10 and 50 total labeled images per dataset, respectively. These results identify two concrete uses of CCD: strengthening an already effective dense predictor and adapting with very few target annotations.

Table R1. Complementary adaptation on frozen GLA-CLIP.

DatasetGLA: 0 target labelsGLA+CCD: 800+200Paired gain [95% CI]
VOC2084.2186.16±0.13+1.95 [+1.11, +2.82]
Context5940.0744.14±0.45+4.07 [+3.77, +4.34]
COCO-Object37.6737.48±0.14-0.19 [-0.53, +0.16]
COCO-Stuff26.9428.14±0.14+1.20 [+1.02, +1.40]
ADE20K20.0223.04±0.13+3.02 [+2.58, +3.49]
Cityscapes40.8347.98±0.39+7.15 [+6.56, +7.74]
Avg641.6244.49±0.06+2.87 [+2.66, +3.07]

Q1. Independent calibration

Yes, in the new mask-only composition protocol. Fitting uses 800 images and candidate selection uses 200 disjoint images within the same 1,000-image total. GLA+CCD achieves 44.49±0.06 Avg6, compared with 41.62 for native GLA, giving +2.87 [+2.66, +3.07]. On ClearCLIP, the corresponding mask-only score is 40.07±0.06, compared with 36.73 for its native reference.

The six-mask candidate pool, margin rule, and +0.50 calibration-acceptance threshold are frozen before evaluation. The test cohort is fixed to all official validation images, and the three primary comparisons use Holm correction. The result tests independent selection for the explicitly defined mask-only composition, with source and split identities recorded.

Q2. Gains on a stronger released baseline

The GLA-CLIP composition supplies this evidence for the tested strong baseline. Keeping its visual encoders, native classifier, vocabulary, geometry, and postprocessing fixed, CCD increases Avg6 by +2.87, from 41.62 to 44.49, with a paired 95% CI of [+2.66, +3.07] and Holm-adjusted p=0.0015. Improvements span VOC +1.95, Context +4.07, Stuff +1.20, ADE +3.02, and City +7.15. Object is −0.19 [−0.53, +0.16]; all six rows appear in Table R1.

GLA itself is training-free with respect to target labels; CCD uses the explicit 800+200 target-image budget. The result establishes complementary target-supervised adaptation on the released strong system, with the gain tied to the channel intervention under a fixed predictor.

Q3. Evidence for the channel-metric intervention

The new experiment isolates mask-only channel adaptation composed with an unchanged dense predictor. GLA+CCD improves five benchmarks, with +2.87±0.06 Avg6 across training seeds and a paired interval [+2.66, +3.07]. Its Context/ADE gains are +4.07/+3.02, alongside the VOC/Stuff/City gains, providing evidence across multiple datasets. Table R1 reports the complete deployed policy, while Table S8 reports the fixed, unselected adapter separately.

The low-label experiment tests a second role for the intervention: at 10/50 total target images, CCD exceeds the same-image LP by +2.93/+5.17 Avg6. These measurements support the channel-metric intervention on strong spatial features and in the extreme-low-label regime.

A1. Reproducible mechanism and evidence

The mechanism is a query-conditioned channel metric with a query-dependent visual normalization. The new composition keeps the base classifier's weights, norms, and biases intact, including GLA's original synonym queries and class mapping. It therefore tests the channel intervention while preserving the existing dense predictor. The +2.87 Avg6 GLA gain and 10/50-image advantages provide its empirical evidence.

The evidence package includes the complete new training, feature-export, and scoring code, fixed fitting and calibration manifests, all candidate decisions, three-seed metrics, and per-image counts for recomputation, so each numerical claim in this response can be reproduced from its implementation.

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

This paper formulates query-conditioned channel reweighting for frozen dense vision-language features, complementing modifications to spatial attention and introduces CCD, which transfers and adapts concept masks and constructs masks directly from competitor errors while retaining dense cosine scoring. In the experimental section, the authors evaluate dataset-level mask selection under an explicit target-label budget and analyze the trade-off between correction and regression across six benchmarks, three feature families, and two backbone sizes.

Critical strengths
  1. The proposed method is described in a detailed and clear manner, with two comprehensive framework figures illustrating its core components.

  2. The experiments are conducted across six standard benchmarks, three feature families, and two backbone sizes.

Critical weaknesses
  1. Unclear motivation. The motivation presented in the Introduction lacks logical clarity, and the supporting evidence is fragmented and insufficient to fully justify the proposed method.

  2. Limited related work. The discussion of related work is limited, with only approximately 20 references cited throughout the paper. This is insufficient to adequately cover existing studies in the field and clearly highlight the novelty of CCD.

  3. Insufficient comparison with sota methods. The most recent method included in the comparison is DIH-CLIP, published at ICCV 2025. Why are more recent methods published at CVPR 2026 or ICML 2026 not included in the experiments? The lack of comparisons with these recent methods makes it difficult to establish the state-of-the-art performance of CCD.

  4. English writing quality. The English writing throughout the manuscript requires substantial improvement in terms of clarity, fluency, and overall academic presentation.

Questions for authors

My main concerns about this paper are detailed in the Critical Weaknesses section, primarily regarding the motivation, related work, claims of state-of-the-art performance, and overall writing quality.

4 comments
AuthorsOpenReview ↗

Motivation, related work, GLA-CLIP comparison and presentation

Dense CLIP segmentation needs both spatially coherent features and class-sensitive comparisons. CCD adapts the channel metric at the comparison stage while keeping the visual backbone and classifier fixed. The new composition experiment establishes this complementarity on GLA-CLIP: Avg6 increases from 41.62 to 44.49 (+2.87 mIoU; paired 95% CI [+2.66, +3.07]), with improvements on five of six datasets. Under extreme annotation scarcity, CCD exceeds a text-initialized linear probe by +2.93 and +5.17 Avg6 at 10 and 50 total labeled images per dataset, respectively. These results identify two concrete uses of CCD: strengthening an already effective dense predictor and adapting with very few target annotations.

W1. Motivation: adapt class comparisons after spatial alignment

Spatially coherent dense features still require a class comparison for each patch. CCD asks whether this comparison should use a query-dependent channel metric. For visual feature f, query t_q, and nonnegative mask m_q,

sq(f)=(f⊙mq)⊤tq∥f⊙mq∥2 ∥tq∥2.s_q(f)=\frac{(f\odot m_q)^\top t_q}{\|f\odot m_q\|_2\,\|t_q\|_2}.

The visual normalization depends on the query, which distinguishes the intervention from an affine class-head update with a shared visual normalization. The decisive experiment holds a strong spatially adapted system fixed: GLA-CLIP 41.62 → GLA+CCD 44.49 Avg6. The 10/50-image results additionally identify where the constrained channel intervention is label-efficient. This gives the motivation a direct chain: dense class comparison → query-conditioned metric → complementary and extreme-low-label gains.

W2. Related-work positioning

The relevant distinction is the intervention stage and supervision setting. SmartCLIP supplies the concept-conditioned prior. ProxyCLIP, NACLIP, ResCLIP, CASS, and CLIPer develop spatial correspondence, attention, or context mechanisms for dense inference. CCD adapts the query-conditioned channel metric of the extracted features, using explicit target-label budgets. Linear probing and text tuning supply the label-matched classifier controls.

The GLA-CLIP composition makes this positioning experimentally testable: its original visual stack and classifier remain frozen, while CCD contributes +2.87 Avg6. This comparison connects the related-work distinction to measured complementarity.

W3. Recent strong systems and complementary adaptation

The updated comparison includes the released GLA-CLIP (CVPR 2026) system. Its reproduced native score is 41.62 Avg6 / 42.42 Avg5. With CCD and 800+200 target images per dataset, the score becomes 44.49±0.06 Avg6 / 45.89±0.05 Avg5. The paired Avg6 improvement is +2.87 [+2.66, +3.07], with five positive per-dataset gains (Table R1).

Table R1. Complementary adaptation on frozen GLA-CLIP.

DatasetGLA: 0 target labelsGLA+CCD: 800+200Paired gain [95% CI]
VOC2084.2186.16±0.13+1.95 [+1.11, +2.82]
Context5940.0744.14±0.45+4.07 [+3.77, +4.34]
COCO-Object37.6737.48±0.14-0.19 [-0.53, +0.16]
COCO-Stuff26.9428.14±0.14+1.20 [+1.02, +1.40]
ADE20K20.0223.04±0.13+3.02 [+2.58, +3.49]
Cityscapes40.8347.98±0.39+7.15 [+6.56, +7.74]
Avg641.6244.49±0.06+2.87 [+2.66, +3.07]

GLA's baseline uses zero target labels; the composition uses the declared target-label budget while retaining its frozen CLIP+DINO encoders, native query vocabulary, classifier, and inference pipeline. This is evidence of added value on a recent strong system.

W4. A result-led introduction and figures

The argument of the paper can be stated in five sentences. Dense vision-language segmentation requires spatially coherent features and class-sensitive comparisons. CCD adapts the query-conditioned channel metric while preserving the visual backbone and classifier. On released GLA-CLIP, this intervention raises Avg6 from 41.62 to 44.49 across six benchmarks. With 10 or 50 total labeled target images per dataset, CCD exceeds a text-initialized linear probe by 2.93 or 5.17 Avg6, respectively. These results establish complementary channel adaptation and a concrete extreme-low-label operating regime.

Table R1 follows the same argument, giving the GLA composition gain with all dataset intervals.

Thank you to the authors for their detailed response, which has partially addressed my concerns. I will therefore keep my score unchanged.

AuthorsOpenReview ↗

A Grateful Follow-up on Any Remaining Concerns

Thank you very much for taking the time to consider our response. We sincerely appreciate your feedback and are grateful that our clarifications have partially addressed your concerns. We fully respect your current assessment.

May we kindly ask whether there are any remaining concerns that are preventing you from considering a higher score? We would greatly appreciate any further guidance and would be happy to provide additional clarification or evidence during the remaining discussion period. We sincerely hope to address your concerns more thoroughly and, if possible, earn your reconsideration of the score.

I am inclined to give a negative score for the following reasons:

  • The revised motivation remains unclear and fails to highlight the novelty of the paper (Reviewers j7gN and c1uU also raised this point). The motivation needs to be supported by systematic quantitative or qualitative experimental data, rather than remaining largely at the narrative level.

  • I pointed out the insufficiency of the related work from my first review: there are only about 26 references, whereas open-vocabulary semantic segmentation is a highly active research area. Such a small number of references makes it difficult to demonstrate the uniqueness of your method. After I emphasized this, I still have not seen any substantial change in this regard.

  • The presentation still has considerable room for improvement, as also noted by Reviewer j7gN. I note that the authors did not submit the latest version of the paper. As a result, I am unable to verify whether my concerns have been addressed, and this does not resolve my concerns.

I appreciate that you have added experiments to demonstrate the advantages of your method. However, the concerns raised in the above three points remain unresolved. Therefore, I maintain my score unchanged.

Reviewer_j7gNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Calibrated Concept Denoising (CCD) for CLIP-based open-vocabulary semantic segmentation. For each text query, CCD applies a non-negative channel mask to frozen visual features before cosine scoring. Masks come from (i) a frozen SmartCLIP mask network, (ii) a residual adapter on the SmartCLIP prior trained with cross-entropy, or (iii) "DirectCDD", a per-class statistic computed on misclassified pixels. With 1,000 densely labeled training images per target dataset, one of seven candidate branches is selected per dataset. With ViT-B/16 under the released ClearCLIP evaluator, the six-dataset average improves over ClearCLIP by +0.48 mIoU (+0.41 with ViT-L/14).

Critical strengths
  • The paper is candid about negative results. It reports the ADE20K regression, correction/regression counts, the gap between fold and validation gains (Table 6), and an earlier validation-informed selection variant (Table 11).
  • The masked cosine score can be computed without materializing a Q×H×W×D tensor (Sec. 3.1).
Critical weaknesses
  1. The gain is negligible relative to the supervision used. The method uses 6,000 densely labeled target images in total, yet the main result is +0.48 Avg6 / +0.49 Avg5 (Table 1). Per-dataset changes are +0.00, +0.10, +0.41, +0.38, −0.23 and +2.20, so Cityscapes accounts for most of the average. All results come from a single seed (seed zero, one epoch) with no variance estimate. Apart from Cityscapes, it is unclear whether any change exceeds run-to-run noise.

  2. The setting is no longer open-vocabulary, and no baseline uses the same labels. Every evaluated class receives target-dataset pixel labels, and the paper states that transfer to withheld classes is not evaluated (Sec. 4.1). CCD is therefore a supervised, per-dataset adaptation of frozen CLIP, yet it is compared only with training-free methods. Several of those methods score higher without any labels: DIH-CLIP reports 42.22 Avg5 in the paper's own Table 1, and ProxyCLIP reports 40.8 on the same five datasets, against 37.96 for CCD. A meaningful evaluation needs baselines that receive the same 1,000 labels, such as a class-wise linear probe or text-embedding tuning on frozen ClearCLIP features. It also needs results on held-out classes.

  3. CCD is a per-dataset selection procedure rather than a single method. Table 2 deploys four different branches across six datasets: unmasked, DirectCDD, P_B and P_ad. They are chosen from seven candidates on the same train1000 split that is used for fitting (Sec. 3.4). The candidate pool itself comes from earlier iterations ("archived v3 and v4 configurations; P_ad corresponds to v5", App. A.1). The pipeline also contains many hand-set constants: a 0.50-mIoU acceptance threshold, a 0.88 cosine gate, 1 + 0.8·minmax scaling, top-200 channels, and a τ_m grid. With this many researcher degrees of freedom and gains of a few tenths of a point, the results are highly exposed to selection effects. Table 6 shows the problem directly: fold gains do not predict validation gains (VOC20 +0.18 → −0.11, Cityscapes +0.28 → −0.43, ADE20K +0.57 → +0.23).

  4. The main learned component hurts most datasets. In Table 7, the prior-preserving adapter produces more regressions than corrections on five of six datasets (e.g., COCO-Object net −36,844). It is deployed only on Cityscapes.

  5. The larger gains are measured against degraded baselines. The +2.86 average (Tables 4, 8, 12) and the feature-family results (Table 3, +2.16 to +4.18) use a "controlled feature-bundle" evaluator. Under it, baselines fall far below their established values:

    • ClearCLIP on Cityscapes: 15.20, vs 30.08 under the released evaluator.
    • ClearCLIP on COCO-Object: 22.46, vs 33.04.
    • SCLIP on Cityscapes: 12.99, vs 32.2 published.

    Gains over such baselines do not support the claimed broader applicability of channel reweighting (Sec. 4.3). The Cityscapes entries also use different checkpoints across tables (24.53 vs 21.21). The 24.53 result comes from a separately trained "dense-mask network" that is not part of the method described in Sec. 3. That network also uses a branch-specific temperature chosen on the training set, and changing τ_m from 1 to 2 alone adds +1.87 on Cityscapes.

  6. Novelty is limited. The transferred-mask branches apply SmartCLIP's concept masks at the patch level. The adapter is a two-layer residual MLP. DirectCDD averages positive competitor–target contrasts per class. The first stated contribution ("formulating query-conditioned channel reweighting") essentially reduces to Eq. 2. Recent training-free methods that address the same dense-feature noise are neither discussed nor compared, e.g., ProxyCLIP, NACLIP, ResCLIP, CASS and CLIPer.

  7. Presentation and reproducibility are poor.

    • Figs. 1–2 repeatedly refer to "certificates", "Certified CDD", "Transfer Certificate" and "Certified Correction → Safe Prediction". The text defines no certificate or guarantee, and it calls the selection "an empirical selection rule" (Sec. 3.4). The figures therefore overstate what the method provides.
    • The method is called CCD in the text but CDD in the figures and component names (DirectCDD, L_CDD).
    • The appendix reads like an experiment log, with internal version IDs v3/v4/v5/v10/v11 and a "policy audit".
Questions for authors
  1. How does CCD compare with simple baselines given the same 1,000 labeled images, such as a class-wise linear probe or text-embedding tuning on frozen ClearCLIP features?
  2. Please report mean ± std over multiple seeds, and over different random draws of the 1,000 training images.
  3. How does CCD perform on classes withheld from the target supervision?
  4. Why do ClearCLIP and SCLIP fall so far below their published values under the controlled evaluator (e.g., Cityscapes 15.20 and 12.99)? What exactly differs from the released evaluator?
  5. What do the "certificates" in Figs. 1–2 refer to? Does the method provide any guarantee?
  6. How was the seven-candidate pool determined? How many configurations were evaluated before it was fixed, including any evaluated on validation data?
3 comments
AuthorsOpenReview ↗

Part 1 of 2: label-matched controls, seed variation and withheld classes

Dense CLIP segmentation needs both spatially coherent features and class-sensitive comparisons. CCD adapts the channel metric at the comparison stage while keeping the visual backbone and classifier fixed. The new composition experiment establishes this complementarity on GLA-CLIP: Avg6 increases from 41.62 to 44.49 (+2.87 mIoU; paired 95% CI [+2.66, +3.07]), with improvements on five of six datasets. Under extreme annotation scarcity, CCD exceeds a text-initialized linear probe by +2.93 and +5.17 Avg6 at 10 and 50 total labeled images per dataset, respectively. These results identify two concrete uses of CCD: strengthening an already effective dense predictor and adapting with very few target annotations.

Tables R1–R2 evaluate the new, explicitly specified mask-only composition protocol. Each 1,000-image budget contains 800 fitting images and 200 disjoint calibration images; the 10/50-image budgets contain 8+2 and 40+10 images. GLA-CLIP itself uses zero target labels, and its CLIP+DINO encoders, native classifier, vocabulary, geometry, and postprocessing remain fixed in the composition comparison. Results use three training seeds on a fixed image draw and all 19,054 official validation task-images. The 2,000 paired image-cluster bootstrap replicates retain shared-image dependence across datasets. The candidate pool and gates were frozen before evaluation.

Table R1. Complementary adaptation on frozen GLA-CLIP.

DatasetGLA: 0 target labelsGLA+CCD: 800+200Paired gain [95% CI]
VOC2084.2186.16±0.13+1.95 [+1.11, +2.82]
Context5940.0744.14±0.45+4.07 [+3.77, +4.34]
COCO-Object37.6737.48±0.14-0.19 [-0.53, +0.16]
COCO-Stuff26.9428.14±0.14+1.20 [+1.02, +1.40]
ADE20K20.0223.04±0.13+3.02 [+2.58, +3.49]
Cityscapes40.8347.98±0.39+7.15 [+6.56, +7.74]
Avg641.6244.49±0.06+2.87 [+2.66, +3.07]

Table R2. Extreme-low-label comparison; same total target-label budget per method.

Total labels per datasetFit + calibrationLP Avg6CCD Avg6CCD−LP [95% CI]
108+234.06±0.9036.99±0.02+2.93 [+2.68, +3.13]
5040+1033.25±0.2538.42±0.13+5.17 [+4.83, +5.41]

Q1. Label-matched controls and the role of CCD

CCD is strongest where labels are scarcest and as a complement to a strong predictor. With 10/50 total images, CCD obtains 36.99/38.42, versus 34.06/33.25 for the same-image linear probe (Table R2). With 800+200 target images, adding CCD to frozen GLA-CLIP raises Avg6 from 41.62 to 44.49, with a paired gain of +2.87 [+2.66, +3.07]. Both methods use the frozen, specified optimization and calibration rules.

The label-matched controls separate classifier adaptation from channel-metric adaptation. At a total budget of 1,000 images with disjoint calibration, the mask-only CCD row is 40.07 Avg6, the text-initialized affine probe is 43.69, and the independently calibrated text-tuning control is 43.84. The probe and text tuning rewrite the classifier and gain capacity as labels grow, whereas CCD keeps the classifier fixed and changes only the per-query channel metric. Because the two act at different stages, CCD can be composed with a strong classifier, as the GLA-CLIP result shows, and it needs far fewer labels to become effective: the linear probe at 50 images remains below CCD at 10.

Q2. Training variation and image draws

The new composition results report three independently fitted training seeds on a fixed image draw. GLA+CCD achieves 44.49±0.06 Avg6; the 10/50-image CCD scores are 36.99±0.02/38.42±0.13. To measure learning variation directly, Table S8 reports a single fixed adapter configuration before candidate selection, margin replacement, or dataset fallback, separately from the deployed policy, so its per-dataset seed deviations come from the fitted adapters themselves.

Image-draw variation is reported for the earlier selection protocol in Table S2. Together, the two tables separate variance from training seeds and from the choice of images.

Q3. Classes withheld from target supervision

The class-withholding diagnostic reports seen/unseen mIoU and their harmonic mean separately (Table S3). It withholds the first five VOC classes and every fourth class in the other datasets; unseen labels are excluded from fitting and calibration. Unseen-class mIoU is 85.62/34.64/34.21/22.09/19.48/31.60 for VOC/Context/Object/Stuff/ADE/City, against 85.62/34.50/34.32/22.07/19.48/33.53 without adaptation. On five of the six datasets, unseen-class performance stays within 0.15 mIoU of the unadapted model, so adapting the channel metric on supervised classes leaves the open-vocabulary behavior on withheld classes intact. All S/U/H entries and supervised controls are retained in Table S3.

AuthorsOpenReview ↗

Part 2 of 2: evaluator alignment, calibration gates, candidate pool and presentation

Q4. Evaluator alignment

The new composition comparison uses the released evaluator and verifies native-baseline parity. Across all six datasets, the recomputed ClearCLIP, GLA-CLIP, and 1,000-image LP scores agree with their reference results. Thus the 41.62→44.49 GLA comparison holds the encoders, classifier, vocabulary, geometry, and postprocessing fixed.

The paired diagnostics isolate the concrete evaluator effects (Table S4). Cityscapes ClearCLIP is 30.07 with aspect-preserving sliding windows and 17.15 with whole-image inference. On COCO-Object, the fixed published rejection threshold 0.4 gives 33.03, compared with 23.32 in the no-rejection ablation. The native-source SCLIP port gives 32.38 on Cityscapes and 33.52 on Object. The low controlled-bundle scores are therefore explained largely by whole-image inference and the absence of background rejection, and the main gain is measured on the released native GLA system.

Q5. Precise meaning of calibration gates

The operational meaning is empirical calibration-gated adaptation. In the new protocol, a candidate is selected using disjoint calibration mIoU and accepted when its calibration improvement reaches 0.50 points. At inference, its patch-score vector replaces the base vector when the top-two class margin is strictly larger. These are explicit, executable acceptance conditions, and their effect is measured on every official validation image.

The mechanism is query-conditioned channel adaptation → disjoint calibration → margin-based score replacement, and "certificate" in the figures refers to passing these two gates.

Q6. Candidate specification and selection

Seven denotes the final archived candidate pool. It combines the unadapted reference, transferred priors, residual adaptation, a competitor-derived mask, selective masking, and top-200 masking; the archived dataset choices are listed in Table S0.

The new claims use a separately specified six-mask candidate pool plus the base fallback: soft/hard priors and adapters at learning rates {1e-4, 1e-3} and epochs {1, 3}. The 0.5 prior threshold, strict class-margin replacement rule, and +0.50-point calibration gate are frozen before evaluation. Fitting uses 800 images and selection uses 200 disjoint images; every official validation image is evaluated with the frozen choices, so no validation access enters candidate selection. Its 44.49 Avg6 GLA composition result is kept distinct from the archived seven-candidate rows.

A1. Adapter deployment and correction–regression diagnostics

The new composition directly tests the channel intervention with the classifier fixed and reports its deployed policy on every dataset. GLA+CCD improves VOC +1.95, Context +4.07, Stuff +1.20, ADE +3.02, and City +7.15; Object is −0.19 [−0.53, +0.16]. Its +2.87 Avg6 therefore reflects gains across five benchmarks.

Correction–regression counts and class-averaged IoU quantify different objectives: the former counts valid pixels equally, so large frequent classes dominate it, whereas mIoU averages class intersection-over-union and rewards gains on rarer classes. All signed net counts are reported in Table S9. Candidate deployment follows the frozen calibration rule, and raw-adapter performance is recorded independently in Table S8.

A2. Contribution relative to concept masks and spatial methods

CCD's contribution is query-conditioned adaptation of the dense comparison metric, supported by a controlled composition experiment. The prior is inherited from SmartCLIP; the residual MLP implements adaptation of that prior. The resulting score changes the query-dependent normalization of the visual feature as well as its weighted dot product.

Spatial feature methods provide the complementary comparison: ProxyCLIP, NACLIP, ResCLIP, CASS, and CLIPer address spatial correspondence, localization, or object context. On the released GLA-CLIP system, keeping its encoders and classifier fixed while applying CCD yields +2.87 Avg6.

A3. Presentation and reproducible evidence

The response uses CCD consistently and organizes the evidence around two questions: complementarity with a strong predictor, and adaptation with very few labels. Tables R1–R2 give the complete six-dataset GLA composition with paired intervals and the explicitly defined 10/50-image regime.

The evidence package records image and split identities, candidate definitions, calibration decisions, per-seed scores, and per-image counts. It supports independent recomputation of every mIoU and interval reported here, and it verifies that the classifier parameters stay frozen throughout fitting.

Thanks for the response. Honestly, it was hard to follow. It brings in a new base model, a new protocol and a new candidate pool, and more importantly, it doesn't directly answer most of my questions. A few specific points:

  1. At 1,000 labels, which is the paper's actual setting, your own numbers show a linear probe (43.69) and text tuning (43.84) both beating CCD (40.07). That was my main concern, and these numbers confirm it. Also, which base model are these numbers on? The response doesn't say.

  2. The low-label comparison doesn't convince me. Your linear probe gets worse going from 10 to 50 labels (34.06 to 33.25), which suggests the baseline isn't set up properly. And CCD at 10 labels (36.99) is basically the unadapted model.

  3. The GLA-CLIP result again compares a model that uses 1,000 labels with one that uses none.

  4. On withheld classes there is no gain, and Cityscapes drops by 1.93. So the method doesn't help classes it wasn't trained on.

  5. Several of my questions are still unanswered: how many configurations you tried before fixing the candidate pool, and whether you still stand behind the original +0.48 result. You now explain the low controlled-evaluator scores by whole-image inference and missing background rejection, yet Tables 3, 4 and 8 stay in the paper. And calling two threshold checks "certificates" is still misleading.

Overall, the new experiments read more like a different paper than a fix to this one. I'm keeping my score.

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

Related papers

Discussion (0)

Sign in to comment.