acceptodds
Under review as a conference paper at ICLR 2027

PartFactor: Recovering Unseen Parts with Decoupled Foreground Prediction

Abstract

Open-vocabulary part segmentation remains challenging: models must distinguish visually similar parts, recover fine spatial boundaries and transfer to categories absent from task supervision. Existing approaches strengthen object–part correspondence and spatial aggregation, but a correctly ranked part can still disappear when recognition uncertainty suppresses foreground. We propose PartFactor, whose object-conditioned decoder first learns conditional part identity by combining structural features, image–text matching and relations among parts. With this decoder fixed, a compact visual head then learns foreground support. A background-logit reparameterization makes foreground probability depend on visual evidence alone, allowing the head to recover missing regions while preserving the decoder’s conditional part distribution. Under the same full-image Pred-All evaluation, PartFactor improves unseen mIoU over a state-of-the-art baseline by 10.12 percentage points on PartImageNet and 23.79 points on PartImageNet-OOD. Fixed-decoder experiments trace the gains to recovery of correctly ranked parts, including at matched background false-positive rates. On PartImageNet, PartFactor needs only about 1/21 as many task-trainable parameters and achieves a 4.54× inference speedup over the same baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
open-vocabulary part segmentation, unseen-part generalization, foreground prediction, vision-language models, dense prediction

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_UVdMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PartFactor, an open-vocabulary part segmentation framework that separates part recognition from foreground prediction. The method first trains an object-conditioned part decoder and then learns a lightweight visual foreground head while keeping the decoder frozen. A background-logit reparameterization is introduced to make foreground probability independent of part scores. Experiments demonstrate substantial improvements in unseen-part mIoU over HOPS, with gains of 10.12 and 23.79 points on PartImageNet and PartImageNet-OOD, respectively.

Critical strengths
  1. The paper identifies foreground rejection as a distinct failure mode in open-vocabulary part segmentation. The motivation is clear, and the proposed probability factorization provides a simple and interpretable way to address this problem.
  2. The experimental results are strong. PartFactor achieves substantial improvements in unseen-part segmentation on both benchmarks while requiring fewer task-trainable parameters and achieving faster inference than HOPS.
  3. The paper provides fairly comprehensive experimental analysis, including fixed-decoder ablations, comparisons at matched background false-positive rates, and detailed error analysis. These experiments help explain how the proposed approach recovers correctly recognized parts that would otherwise be classified as background.
Critical weaknesses
  1. The methodological novelty appears somewhat limited. Separating foreground probability from conditional class predictions is a relatively straightforward idea, and the proposed background-logit reparameterization follows naturally from this factorization. The paper could better clarify what distinguishes this formulation from existing approaches that model foreground and class predictions separately.
  2. Although the fixed-decoder experiments support the effectiveness of foreground decoupling, it remains difficult to determine how much this mechanism contributes to the overall improvements over HOPS. PartFactor also uses different pretrained encoders, a new decoder architecture, and a different training pipeline. These differences make it difficult to attribute the large performance gains specifically to the proposed decoupling strategy.
  3. The effectiveness of the approach appears to depend on the underlying decoder. Applying foreground adaptation to HOPS and PartCATSeg leads to only marginal improvements, which raises questions about its generalizability. In addition, the gains are mainly concentrated in larger parts, while performance on small parts remains limited.
Questions for authors
  1. Could the authors provide additional architecture-level ablations or comparisons using matched pretrained encoders to better isolate the contribution of foreground decoupling from other architectural changes?
  2. The experiments on HOPS and PartCATSeg suggest that foreground adaptation is not consistently beneficial across different decoders. Could the authors further clarify under what conditions the proposed approach is expected to be effective, and whether its effectiveness can be predicted from the initial decoder's foreground rejection behavior?
  3. On PartImageNet-OOD, the original-background visual-plus-score head already achieves 55.42% unseen mIoU, compared with 56.11% for the proposed score-decoupled head. This relatively small difference suggests that much of the improvement may come from introducing spatial visual foreground modeling rather than score decoupling itself. Could the authors further clarify the relative contributions of these two components?

If the authors can address most of my concerns, I would be willing to raise my score.

7 comments
AuthorsOpenReview ↗

Response to Reviewer UVdM (1/5)

We sincerely thank you for the thoughtful review and for recognizing the meaningful failure diagnosis, strong empirical results, and controlled analysis of PartFactor. We especially appreciate your questions about methodological novelty, the contribution of pretrained representations, and the conditions under which foreground adaptation is useful.

We have incorporated your feedback and marked the key revisions addressing your comments in orange\color{#B50}\textbf{orange} in the revised manuscript.

These questions go directly to what we wanted to understand: when a model misses a part, how much of the correct prediction is already present, and what is the simplest intervention that can recover it?

The revised manuscript now includes experiments with normally background-supervised decoders, matched pretrained initializations, independently retrained architecture controls, and additional tests of foreground evidence and score suppression. These experiments provide a clearer account of both the benefits and the limits of foreground learning.

Critical Weaknesses\color{#05B}\textbf{Critical Weaknesses}

Weakness W1. Limited methodological novelty\color{#B50}\textbf{Weakness W1. Limited methodological novelty}

W1.1. The simplicity of the factorization and reparameterization\color{#B50}\textbf{W1.1. The simplicity of the factorization and reparameterization}

Original comment:

The methodological novelty appears somewhat limited. Separating foreground probability from conditional class predictions is a relatively straightforward idea, and the proposed background-logit reparameterization follows naturally from this factorization.

Response\color{#073}\textbf{Response}: PartFactor uses a simple mathematical construction to identify a measurable source of segmentation error and turn it into a controlled learning problem.

The underlying probability factorization is standard. Its usefulness here comes from separating two errors that produce the same visible outcome.

A pixel may disappear because the decoder prefers the wrong part identity. Alternatively, it may prefer the correct identity yet lose the final comparison with background.

Equation (3) shows why this happens, and Equation (4) defines the event as correct-part rejection. Measuring it tells us which correct predictions are already available but discarded.

We then design a direct intervention: freeze the decoder, preserve the conditional part distribution on its prediction grid, and learn only foreground support. This makes it possible to measure how much segmentation performance can be recovered without relearning part identity.

The experiment remains meaningful after ordinary background supervision. On the joint-trained PI decoder, conditional unseen-part accuracy remains 85.90%, while correct-part rejection falls from 18.94% to 17.10% after visual foreground adaptation (Appendix Table 27).

The mathematical simplicity is important to the experimental design. By changing only the foreground decision, we can identify what that decision contributes without introducing further changes to part recognition.

W1.2. Distinction from existing approaches\color{#B50}\textbf{W1.2. Distinction from existing approaches}

Original comment:

The paper could better clarify what distinguishes this formulation from existing approaches that model foreground and class predictions separately.

Response\color{#073}\textbf{Response}: Our contribution connects a precisely measured failure mode, a controlled recovery experiment, and an economical intervention whose effects and limits can be established.

The matched-background-error experiments provide a further test. At a 2% background-FPR budget, correct unseen-part recall increases from \color{#073}\textbf{62.19\\% to 70.95\\% on PI} and \color{#073}\textbf{74.86\\% to 80.60\\% on OOD} (Section 5.3 and Appendix P).

Thus, the learned foreground model improves the spatial selection of correctly named parts, beyond shifting a single global threshold.

The intervention is deliberately small. The foreground head reuses existing features and trains only 22,177 parameters\color{#073}\textbf{22,177 parameters} on PI. This allows us to investigate the source of an error without making the recognition model more complicated.

The revised Related Work also discusses the relevant approaches to separating recognition, spatial grouping, and objectness, including ZegFormer, MaskCLIP, SAN, FC-CLIP, and OVRCOAT.

The resulting contribution connects a specific failure mode, a measurable recovery opportunity, and a computationally inexpensive intervention. It provides a way to decide when additional capacity should improve recognition and when it should improve foreground support.

AuthorsOpenReview ↗

Response to Reviewer UVdM (2/5)

Weakness W2. Attribution of the performance improvements\color{#B50}\textbf{Weakness W2. Attribution of the performance improvements}

W2.1. The independent contribution of foreground learning\color{#B50}\textbf{W2.1. The independent contribution of foreground learning}

Original comment:

Although the fixed-decoder experiments support the effectiveness of foreground decoupling, it remains difficult to determine how much this mechanism contributes to the overall improvements over HOPS.

Response\color{#073}\textbf{Response}: We now measure the foreground contribution directly on fixed, background-supervised decoders and distinguish it from complete-system performance.

The revised default trains each decoder with part and background supervision from its first update. We then freeze the decoder and learn foreground support.

Across the four benchmarks, unseen mIoU improves as follows:

BenchmarkJoint training onlyPartFactorForeground gain
PartImageNet48.6851.52+2.83
PartImageNet-OOD53.4856.88+3.40
Pascal-Part-11622.4732.24+9.77
ADE20K-Part-2347.3611.24+3.88

Each pair shares the same decoder and conditional part distribution on the decoder grid. Gains are computed from unrounded mIoU values.

The PI and OOD gains have paired-image 95% intervals of [1.76, 4.65] and [3.16, 3.66] points. These measure evaluation-image variation with trained model weights fixed.

We further examine whether the foreground contribution persists across different learned representations. Three complete PI decoder trainings with independent initializations yield unseen-mIoU improvements of 2.83, 5.59, and 3.58 points after foreground adaptation, averaging 4.00 ± 1.42 points\color{#073}\textbf{4.00 ± 1.42 points} (Appendix S, Table 42). Separately, three foreground-head initializations on the fixed default decoder reach 51.77 ± 0.27% unseen mIoU on PI and 56.80 ± 0.07% on OOD (Appendix Table 36). These experiments distinguish variability in the learned decoder from variability in the foreground intervention itself. In both settings, the foreground gain remains positive.

These paired experiments measure the contribution of the foreground stage without changing the decoder's conditional part representation.

W2.2. Differences in pretrained encoders, decoder architecture,\color{#B50}\textbf{W2.2. Differences in pretrained encoders, decoder architecture,} and training pipeline\color{#B50}\textbf{and training pipeline}

Original comment:

PartFactor also uses different pretrained encoders, a new decoder architecture, and a different training pipeline. These differences make it difficult to attribute the large performance gains specifically to the proposed decoupling strategy.

Response\color{#073}\textbf{Response}: We investigate representation quality through separate controls and use fixed-decoder experiments to isolate foreground adaptation from those differences.

To examine the effect of pretrained representations, we replace the default encoders with HOPS's DINOv2-S/14 and CAT-Seg-initialized CLIP. Foreground adaptation still improves unseen PI mIoU from 39.44% to 41.29% on the same frozen decoder (Appendix Table 29).

Separately, the architecture ablations reveal where the representation gains come from. Removing the DINO structural pathway reduces conditional unseen-part accuracy from 87.82% to 71.97%. Removing object context causes a much larger performance loss than simplifying the decoder's iterative refinement (Appendix Tables 31 and 32).

These experiments distinguish two contributions.

The structural and semantic representations determine which part identities the model can recognize. The foreground stage determines which of those identities survive the final prediction.

The complete-system comparison with HOPS includes both effects, while the paired fixed-decoder experiments isolate the foreground intervention. This distinction is now explicit in the revised manuscript.

Together, W2.1 and W2.2 provide a more precise attribution of PartFactor's performance: recognition quality establishes the available part identities, and foreground specialization contributes additional recovery while retaining the trained part decoder.

Weakness W3. Generalization across decoders and part sizes\color{#B50}\textbf{Weakness W3. Generalization across decoders and part sizes}

W3.1. Dependence on the underlying decoder\color{#B50}\textbf{W3.1. Dependence on the underlying decoder}

Original comment:

The effectiveness of the approach appears to depend on the underlying decoder. Applying foreground adaptation to HOPS and PartCATSeg leads to only marginal improvements, which raises questions about its generalizability.

Response\color{#073}\textbf{Response}: The revised experiments show that foreground recovery depends on the error structure of the starting decoder.

AuthorsOpenReview ↗

Response to Reviewer UVdM (3/5)

A foreground head can recover correctly ranked parts that are rejected as background. Its effect is therefore determined partly by how many such predictions exist and partly by the errors introduced when they are accepted.

This helps explain the external-decoder results.

On OOD, retrained HOPS has zero measured correct-part rejection at its original operating point (Appendix Table 27). A new foreground head has little of this particular error to recover. A source-calibrated global bias reaches 32.65% unseen mIoU, slightly above the spatial-head alternatives (Appendix Table 22).

PartCATSeg presents a different case. Its fixed 20,000-update checkpoint has 6.08% correct-part rejection on PI (Appendix Table 27). The separate foreground-head experiment uses the earlier 8,000-update decoder, with unseen mIoU of 35.69% before adaptation and 35.15–36.04% after (Appendix Table 23). These experiments show that having correct identities available for recovery does not by itself ensure a net improvement. The learned foreground decision must also avoid losing existing correct labels or introducing new errors.

These observations establish an important distinction: correct-part rejection identifies a recovery opportunity, while the eventual performance gain depends on the selectivity of the learned foreground decision.

The external decoder experiments therefore help explain when targeted foreground learning is useful and why its effect varies across models.

W3.2. Recovery of small parts\color{#B50}\textbf{W3.2. Recovery of small parts}

Original comment:

In addition, the gains are mainly concentrated in larger parts, while performance on small parts remains limited.

Response\color{#073}\textbf{Response}: We have repeated the area analysis using the revised background-supervised models and find measurable improvements across part sizes, including small regions.

On PI, mean image–part IoU for regions smaller than 1,000 pixels rises from 9.21% to 12.04%. On OOD, 1,086 such regions improve from 15.04% to 19.83%, with median IoU rising from 3.01% to 11.93% (Appendix Table 33).

The smallest PI regions retain a zero median, indicating that fine boundaries and ambiguous part identities remain difficult. However, foreground specialization produces measurable improvements even in this challenging size range.

The external decoder and part-size experiments therefore clarify when targeted foreground learning is useful. They also identify where better recognition and fine-scale evidence are the more appropriate next steps.

Questions for Authors\color{#05B}\textbf{Questions for Authors}

Question Q1. Architecture-level ablations and matched pretrained encoders\color{#B50}\textbf{Question Q1. Architecture-level ablations and matched pretrained encoders}

Original question:

Could the authors provide additional architecture-level ablations or comparisons using matched pretrained encoders to better isolate the contribution of foreground decoupling from other architectural changes?

Response\color{#073}\textbf{Response}: We have added both the matched-initialization control and independently retrained architectural ablations.

First, we replace PartFactor's default pretrained encoders with HOPS's DINOv2-S/14 and CAT-Seg-initialized CLIP, removing DINOv3-L, SigLIP 2, and ResNet-50.

The jointly trained PartFactor decoder reaches 39.44% unseen PI mIoU, increasing to 41.29% after foreground adaptation with the decoder fixed (Appendix Table 29).

Second, we retrain architecture replacements under a common PI schedule (Appendix Table 31).

The full architecture reaches 47.06% unseen mIoU. Removing object context reduces unseen mIoU to 0.00%, while replacing the three refinement rounds with a single image–text matching step retains 46.83%.

This result suggests that object-conditioned evidence is especially important for recognition, whereas additional iterative refinement contributes relatively little under the tested schedule.

The structural-feature ablation provides another view. Removing DINOv3 reduces conditional unseen-part accuracy from 87.82% to 71.97%, accompanied by an unseen-mIoU decrease from 47.06% to 30.05% (Appendix Table 32).

These controls reveal which parts of the representation supply useful part identities. The matched-decoder foreground experiment then measures how those identities can be retained without relearning the representation.

Reviewer_e7UpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets open-vocabulary part segmentation and observes that a correctly ranked part can still lose to background because foreground probability is tied to part scores. PartFactor trains an object-conditioned part decoder on frozen DINOv3-L, CLIP, SigLIP 2 and ResNet-50 features, then freezes it and trains a small visual foreground head with a background-logit reparameterization that makes foreground depend on visual features alone. Under the authors’ full-image protocol with retrained baselines, unseen mIoU rises over HOPS from 35.18 to 45.30 on PartImageNet and from 32.32 to 56.11 on PartImageNet-OOD, with 2.72M task-trainable parameters and a 4.54× inference speedup.

Critical strengths
  • Identifies a precise failure mode, correctly ranked parts lost to background, formalizes it with an exact factorization and a measurable rejection rate, and fixes it with a 22K-parameter visual head plus a one-line background-logit reparameterization.
  • The controlled analysis is thorough, with fixed-decoder comparisons, a factorial study of inputs and parameterization, three seeds, paired confidence intervals and matched false-positive-rate recall curves.
  • Reports unfavorable evidence openly, including published baseline numbers under their original protocols and the null result on HOPS and PartCATSeg, and delivers a real efficiency gain of 4.54× inference speed with 11 minutes of task training.
Critical weaknesses
  • The motivating failure is self-inflicted and the simplest fix is untested. The introduction presents correctly ranked parts disappearing into background as a general weakness of existing methods, yet Section 4.5 finds it absent in HOPS and the proposed head gives PartCATSeg no gain. The failure is large in the proposed decoder because its first stage gives background no positive examples. Training the decoder in one stage with background supervision is never reported, so the paper cannot show the two-stage factorization adds anything beyond supervising background.

  • The headline comparison exists only under the authors’ own protocol. Under HOPS’s published protocol HOPS reaches 57.46 on PI and 45.08 on OOD, against 35.18 and 32.32 under the paper’s protocol with author-retrained baselines. PartFactor is never evaluated under the original protocol.

  • The gain is confounded with stronger encoders. PartFactor fuses DINOv3-L, CLIP, SigLIP 2 and a ResNet-50, HOPS uses CLIP and an older DINO, and no encoder-matched control is given. The proposed decoder with a score-coupled head scores 35.59, level with HOPS, and the head has no effect on HOPS, so the source of the gain is recognition quality rather than the mechanism. The 1/21 parameter claim omits an 846M-parameter inference model.

  • Recovery is uneven. Parts below one thousand pixels gain 0.40 points, regions above twenty thousand pixels supply three quarters of the gain, and wrong-part assignments rise from 8.92 to 12.75 percent.

Questions for authors
  1. No code, checkpoints or supplementary material are provided. Will the code, the trained decoders, the retrained baseline checkpoints and the evaluation scripts be released?
  2. What does PartFactor score under HOPS’s original protocol, and do the retrained HOPS and PartCATSeg reproduce their published numbers under that protocol? Which protocol differences account for the drop from 57.46 to 35.18?
  3. What does PartFactor achieve with HOPS’s encoders, or HOPS with DINOv3-L, so that the contribution of the encoders can be separated from the foreground head?
  4. What does the decoder achieve when trained in a single stage with background supervision, without the second-stage head?
  5. What is the correct-part rejection rate of PartCATSeg and of a decoder trained with background supervision, so that the prevalence of the motivating failure can be judged beyond the proposed decoder?
8 comments
AuthorsOpenReview ↗

Response to Reviewer e7Up (1/6)

We sincerely thank you for the detailed review and for recognizing the precise failure mode, thorough controlled analysis, and practical efficiency of PartFactor. Your questions about background supervision, encoder contributions, and evaluation protocols were particularly valuable. They led us to revisit the simplest possible alternative to our original method: what if the decoder learns background normally from the beginning? Would separately learning foreground still offer any benefit?

We have incorporated your feedback and marked the key revisions addressing your comments in green\color{#073}\textbf{green} in the revised manuscript.

We have now tested this alternative directly and made joint part–background training the default. The revised paper also separates the contribution of foreground learning from the representations and computational choices of the complete system.

Critical Weaknesses\color{#05B}\textbf{Critical Weaknesses}

Weakness W1. The motivating failure and the single-stage alternative\color{#B50}\textbf{Weakness W1. The motivating failure and the single-stage alternative}

W1.1. Is correct-part rejection specific to the proposed decoder?\color{#B50}\textbf{W1.1. Is correct-part rejection specific to the proposed decoder?}

Original comment:

The motivating failure is self-inflicted and the simplest fix is untested. The introduction presents correctly ranked parts disappearing into background as a general weakness of existing methods, yet Section 4.5 finds it absent in HOPS and the proposed head gives PartCATSeg no gain.

Response\color{#073}\textbf{Response}: Correct-part rejection is measurable across several decoders, but its prevalence and the benefit of recovery depend on their existing error patterns.

The revised Appendix Table 27 measures the event on the original-resolution predictions of the evaluated decoders.

On OOD, retrained HOPS has zero correct-part rejection\color{#B50}\text{zero correct-part rejection} at its original operating point. It therefore has little of this particular error to recover. PartCATSeg has nonzero rejection, including 6.08% on PI and 3.51% on OOD, but additional foreground acceptance can introduce wrong-part assignments and background false positives.

The distinction is important. Correct-part rejection identifies correctly recognized pixels currently assigned to background. It provides a measurable opportunity for recovery, while the net gain depends on how accurately the foreground decision uses that opportunity.

This explains why foreground adaptation benefits different decoders differently. It also motivates measuring recognition and rejection separately instead of treating every missing part as the same error.

W1.2. Does the failure remain after positive background supervision?\color{#B50}\textbf{W1.2. Does the failure remain after positive background supervision?}

Original comment:

The failure is large in the proposed decoder because its first stage gives background no positive examples.

Response\color{#073}\textbf{Response}: The revised default receives positive background supervision from its first update, yet correct-part rejection remains measurable and foreground adaptation still improves segmentation.

The revised default first trains an object-conditioned decoder using both part and background labels. Once this decoder is fully trained, we freeze it and learn foreground support with a small visual head.

On PI, the jointly trained decoder reaches 48.68% unseen mIoU. Foreground adaptation raises this to \color{#073}\textbf{51.52\\%} without changing the decoder. The same experiment gives 53.48% to 56.88% on OOD, 22.47% to 32.24% on Pascal, and 7.36% to 11.24% on ADE (Tables 1 and 3).

The crucial property of these comparisons is that the conditional part distributions remain fixed on the decoder grid. The additional gains come from learning a different foreground decision on an already trained representation.

The revised PI decoder also retains 18.94% correct-part rejection on its prediction grid before foreground adaptation. The learned visual head reduces this to 17.10%, while conditional unseen-part accuracy remains 85.90% (Appendix Table 27).

This establishes that foreground specialization addresses a measurable remaining error after ordinary background learning.

W1.3. Is single-stage background training sufficient, or does\color{#B50}\textbf{W1.3. Is single-stage background training sufficient, or does} the second stage add something?\color{#B50}\textbf{the second stage add something?}

Original comment:

Training the decoder in one stage with background supervision is never reported, so the paper cannot show the two-stage factorization adds anything beyond supervising background.

Response\color{#073}\textbf{Response}: We have tested single-stage background-supervised training directly and compared it with foreground specialization under an equal training budget.

AuthorsOpenReview ↗

Response to Reviewer e7Up (2/6)

On PI, continuing joint training for 95,500 image presentations produces 48.68% unseen mIoU. Dividing the same budget between joint training and foreground specialization produces 49.88%, with harmonic mIoU changing from 51.70% to 51.58% (Table 2).

Completing the full joint decoder training before adding foreground adaptation gives the default result of 51.52% unseen mIoU with 111,000 total presentations.

We assess the result at two levels. With the joint-trained decoder fixed, three foreground-head initializations reach 51.77 ± 0.27% unseen mIoU on PI and 56.80 ± 0.07% on OOD, with every initialization improving its starting decoder (Appendix Table 36). We also repeat the complete PI decoder training from three different initializations, followed by foreground adaptation on each frozen decoder. The corresponding unseen-mIoU gains are 2.83, 5.59, and 3.58 points, averaging 4.00 ± 1.42 points\color{#073}\textbf{4.00 ± 1.42 points} (Appendix Table 42). Thus, foreground specialization remains beneficial when the decoder itself is retrained, beyond variation in the foreground head alone.

We also compare against a simpler alternative: selecting one global background bias on source validation. This gives 46.38% on PI and 53.34% on OOD, below the learned spatial heads.

Together, W1.1–W1.3 establish that foreground specialization has measurable value after normal background learning, and that spatial selection provides more than a uniform adjustment to the background score.

Weakness W2. Evaluation protocol and published results\color{#B50}\textbf{Weakness W2. Evaluation protocol and published results}

W2.1. Differences between the published HOPS results and locally\color{#B50}\textbf{W2.1. Differences between the published HOPS results and locally} retrained baselines\color{#B50}\textbf{retrained baselines}

Original comment:

The headline comparison exists only under the authors' own protocol. Under HOPS's published protocol HOPS reaches 57.46 on PI and 45.08 on OOD, against 35.18 and 32.32 under the paper's protocol with author-retrained baselines.

Response\color{#073}\textbf{Response}: We have applied the released PI scoring functions to the saved predictions and now report the different evaluation settings separately.

The primary comparison evaluates all models on the same complete images using fixed-class Pred-All, with zero-IoU classes retained. This ensures that every candidate part and background prediction is evaluated under the same metric definition.

Appendix Table 34 also applies the released PI scoring functions to the same saved predictions. These calculations distinguish fixed-class evaluation from positive-IoU averaging and from an additional setting that supplies ground-truth background.

For retrained HOPS, the corresponding PI unseen mIoU values are 35.10%, 41.08%, and 43.82% using the fixed 20,000-update checkpoint. The published HOPS result is 57.46%. Switching from fixed-class to positive-IoU averaging increases our measured score by 5.98 points. The remaining 16.38-point difference\color{#B50}\textbf{16.38-point difference} from the published result is documented alongside the training and implementation differences in Appendix N. We report these evaluation settings separately and do not attribute the remaining difference to any individual factor without a controlled measurement.

To make the comparison inspectable, Appendix N records the training configurations, checkpoint choices, evaluation rules, and implementation details. Appendix R separately retains the published results alongside our common-protocol measurements.

W2.2. PartFactor under the released evaluation rules\color{#B50}\textbf{W2.2. PartFactor under the released evaluation rules}

Original comment:

PartFactor is never evaluated under the original protocol.

Response\color{#073}\textbf{Response}: We now report PartFactor under the released PI scoring rules, alongside every locally evaluated baseline.

For PartFactor, unseen mIoU is 51.52% under fixed-class Pred-All, 54.49% under positive-IoU class averaging, and 58.74% when ground-truth background is supplied (Appendix Table 34).

These calculations answer different evaluation questions. Positive-IoU averaging changes the classes included in the mean. Supplying ground-truth background additionally replaces part of the model's prediction with oracle information. We report these conditions separately.

Our comparisons also include results in which other methods outperform PartFactor. On Pascal, retrained PartCATSeg achieves 36.63% unseen mIoU, compared with PartFactor's 32.24%. Under direct PI-to-ADE transfer, CAT-Seg reaches 1.78% All mIoU, compared with PartFactor's 0.96%.

The useful distinction is between the two questions these experiments answer. Cross-method comparisons assess complete systems under the recorded conditions. The fixed-decoder experiments isolate foreground adaptation without changing the recognition model or evaluation.

AuthorsOpenReview ↗

Response to Reviewer e7Up (3/6)

The resulting evidence supports a clear conclusion: foreground learning contributes to PartFactor independently of the performance differences among complete OVPS architectures.

Weakness W3. Encoder contributions, mechanism attribution, and\color{#B50}\textbf{Weakness W3. Encoder contributions, mechanism attribution, and} efficiency accounting\color{#B50}\textbf{efficiency accounting}

W3.1. Stronger encoders and matched pretrained features\color{#B50}\textbf{W3.1. Stronger encoders and matched pretrained features}

Original comment:

The gain is confounded with stronger encoders. PartFactor fuses DINOv3-L, CLIP, SigLIP 2 and a ResNet-50, HOPS uses CLIP and an older DINO, and no encoder-matched control is given.

Response\color{#073}\textbf{Response}: We have quantified the contribution of pretrained representations and added a control using HOPS's pretrained initialization resources.

PartFactor combines structural, semantic, and object-level evidence for part recognition. These representations are important for the quality of the conditional part distribution, as shown by the encoder ablations in Appendix Tables 30 and 32.

To examine the foreground mechanism without relying on the default pretrained resources, we replace them with HOPS's DINOv2-S/14 and CAT-Seg-initialized CLIP. SigLIP 2 and ResNet-50 are removed.

Under this initialization, the joint-trained PartFactor decoder reaches 39.44% unseen PI mIoU. Adapting foreground on the same decoder raises it to \color{#073}\textbf{41.29\\%, a gain of 1.85 points} (Appendix Table 29).

The encoders are frozen in PartFactor and fine-tuned in the retrained HOPS implementation, so the two complete systems retain different training policies and architectures. The paired 1.85-point improvement directly measures the foreground contribution under the alternative pretrained initialization.

W3.2. Recognition quality versus foreground learning\color{#B50}\textbf{W3.2. Recognition quality versus foreground learning}

Original comment:

The proposed decoder with a score-coupled head scores 35.59, level with HOPS, and the head has no effect on HOPS, so the source of the gain is recognition quality rather than the mechanism.

Response\color{#073}\textbf{Response}: The revised fixed-decoder experiments measure the foreground contribution directly, separately from changes in recognition quality.

The 35.59% score-coupled result belongs to the earlier part-only-pretrained decoder and one foreground configuration. The revised default instead trains the decoder with ordinary part–background supervision before foreground adaptation.

With this default representation, foreground adaptation adds 2.83 unseen-mIoU points on PI\color{#073}\textbf{2.83 unseen-mIoU points on PI} and 3.40 points on OOD\color{#073}\textbf{3.40 points on OOD} to already background-supervised decoders.

This distinguishes two sources of improvement. Better representations provide more correct part identities. Foreground adaptation improves which of those identities appear in the final segmentation.

The complete-system comparison includes both effects. The paired fixed-decoder experiments isolate the contribution of foreground learning.

W3.3. Task-trainable parameters versus complete model size\color{#B50}\textbf{W3.3. Task-trainable parameters versus complete model size}

Original comment:

The 1/21 parameter claim omits an 846M-parameter inference model.

Response\color{#073}\textbf{Response}: The revised manuscript explicitly distinguishes task-trainable parameters, additional foreground-head parameters, and total deployed model size.

The PI foreground head trains 22,177 parameters\color{#073}\textbf{22,177 parameters}, while the complete two-stage task model trains 2.72M. Retrained HOPS trains 56.83M task parameters, approximately 21 times as many.

After encoding the fixed vocabulary, PartFactor's inference model, including the frozen visual encoders, contains 846.32M parameters. Its training efficiency comes from learning a compact task model on top of existing representations, not from removing the pretrained encoders.

The efficiency advantage also holds under a more tightly controlled inference comparison. With both models caching the fixed vocabulary, receiving 16 warm-up calls, and using the same memory-reuse and timing procedure, HOPS averages 283.18 ± 12.20 ms/image, while PartFactor averages 46.73 ± 0.88 ms/image, giving a 6.06× speedup\color{#073}\textbf{6.06× speedup} on the same 286 PI images (Appendix S, Table 41). The original measurements using the recorded inference pipelines yield 415.06 and 45.01 ms/image, respectively, corresponding to 9.22× (Table 5). Both comparisons demonstrate a substantial practical latency advantage, with the controlled experiment separating it more clearly from caching and measurement differences.

Together, W3.1–W3.3 show the practical value of the design: useful part recognition can be retained and improved through a small additional learning task, without repeatedly training large recognition components.

Reviewer_ff9LOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2(+1)
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper proposes PartFactor, an open-vocabulary part segmentation framework that separates two prediction problems: identifying which part a pixel belongs to and determining whether that pixel belongs to foreground.

Critical strengths
  • Clear problem formulation and motivation. The paper addresses a concrete and practically relevant problem. In particular, the distinction between incorrect part identification and correctly ranked parts being rejected as background is meaningful and provides a reasonable motivation for the proposed method.

  • Transparent failure analysis. The authors provide a relatively thorough analysis of the method's limitations, including cases where foreground recovery increases incorrect part assignments, degrades performance on individual images, or fails to improve other part decoders. Such analyses strengthen the credibility of the experimental findings.

  • Simplicity of the proposed method. The method is conceptually simple and straightforward to implement, potentially facilitating its integration into existing segmentation frameworks.

Critical weaknesses

Overall, I find the technical contribution incremental and the experimental evaluation insufficient for an ICLR publication. In its current form, the work resembles a preliminary proof-of-concept rather than a sufficiently developed research contribution. Substantial additional experiments and more comprehensive comparisons with state-of-the-art methods are needed to support the paper's claims.

  1. Incomplete benchmark coverage. Two major benchmarks for open-vocabulary part segmentation, Pascal-Part-116 and ADE20K-Part-234, are missing. Both are widely adopted by several state-of-the-art methods cited in this paper. Evaluating exclusively on PartImageNet and PartImageNet-OOD significantly limits the comprehensiveness of the experimental evaluation.

  2. Insufficient evidence of cross-dataset generalization and potentially unfair comparisons. The paper does not evaluate cross-dataset transfer beyond the PartImageNet dataset family, despite emphasizing generalization to unseen parts. Cross-dataset evaluation is particularly important for establishing the transferability of open-vocabulary segmentation methods. Furthermore, PartFactor leverages multiple powerful pretrained feature extractors, including DINOv3-L, CLIP, SigLIP 2, and an ImageNet-supervised ResNet-50. Without controlling for differences in pretrained representations, it is difficult to determine whether the reported improvements arise from the proposed method or from the stronger combination of pretrained features.

  3. Limited methodological novelty. The central formulation decomposes the segmentation distribution into a foreground probability and a conditional distribution over part identities. The primary technical contribution is a background-logit reparameterization that makes foreground probability independent of part logits, conditioned on visual features. While mathematically sound, this is essentially a standard probabilistic factorization, and the reparameterization is equivalent to introducing an independent sigmoid foreground predictor. The paper does not sufficiently establish how this formulation differs from, or advances beyond, existing approaches that separate class-agnostic localization from semantic recognition.

  4. Insufficient discussion of closely related methodological precedents. Although the related-work section covers several recent open-vocabulary part segmentation methods, it overlooks important prior work on decoupling class-agnostic segmentation and semantic classification. These studies are directly relevant to the proposed formulation and should be discussed to better contextualize the paper's novelty.

Questions for authors
  1. Can the authors ablate DINOv3-L, SigLIP 2, and the ImageNet-trained ResNet-50, reporting both the conditional decoder accuracy and final unseen mIoU? How much improvement remains when matching pretrained visual features with the baselines?
6 comments
AuthorsOpenReview ↗

Response to Reviewer ff9L (1/4)

We sincerely thank you for the careful review and for recognizing the clear problem formulation, transparent failure analysis, and simplicity of PartFactor. We particularly appreciate your questions about experimental coverage, pretrained representations, and methodological novelty. They encouraged us to examine the question that first motivated this work more deeply: when a part disappears from a segmentation, how much of the correct prediction is already present in the model, and how little additional learning is needed to recover it?

We have incorporated your feedback and marked the key revisions addressing your comments in blue\color{#05B}\textbf{blue} in the revised manuscript.

The revised manuscript now includes four benchmarks, direct cross-dataset transfer, ordinary background supervision, matched training budgets, encoder and architecture controls, and additional foreground experiments. We discuss how these results strengthen the scientific contribution below.

Critical Weaknesses\color{#05B}\textbf{Critical Weaknesses}

Weakness W1. Incomplete benchmark coverage\color{#B50}\textbf{Weakness W1. Incomplete benchmark coverage}

Original comment:

Incomplete benchmark coverage. Two major benchmarks for open-vocabulary part segmentation, Pascal-Part-116 and ADE20K-Part-234, are missing. Both are widely adopted by several state-of-the-art methods cited in this paper. Evaluating exclusively on PartImageNet and PartImageNet-OOD significantly limits the comprehensiveness of the experimental evaluation.

Response\color{#073}\textbf{Response}: We have extended the evaluation to both Pascal-Part-116 and ADE20K-Part-234, including complete-system comparisons and controlled foreground adaptation.

The new experiments cover all 850 Pascal images and 1,615 ADE images. Each model is trained on its benchmark's source data and evaluated using the same original-resolution, fixed-class Pred-All metric.

PartFactor achieves \color{#073}\textbf{32.24\\% unseen mIoU on Pascal} and \color{#073}\textbf{11.24\\% on ADE}. Retrained PartCATSeg reaches 36.63% and 11.10%, respectively. Table 3 presents the common evaluation, while Appendix R retains the complete comparisons with published methods and official checkpoints.

Beyond evaluating the complete systems, we also ask whether the foreground intervention itself remains useful on these more complex datasets. With the part decoder fixed, foreground adaptation improves unseen mIoU from 22.47% to 32.24% on Pascal and 7.36% to 11.24% on ADE, gains of 9.77 and 3.88 points.

These experiments reveal a useful distinction. Foreground adaptation provides substantial improvements in multi-object scenes, but its value still depends on the part identities supplied by the decoder. Pascal retains stronger conditional part recognition than ADE, and the additional analysis identifies where better semantic correspondence remains necessary.

The revised evaluation therefore establishes both the broader applicability of foreground learning and the limits imposed by part recognition.

Weakness W2. Cross-dataset generalization and potentially unfair comparisons\color{#B50}\textbf{Weakness W2. Cross-dataset generalization and potentially unfair comparisons}

W2.1. Cross-dataset transfer\color{#B50}\textbf{W2.1. Cross-dataset transfer}

Original comment:

Insufficient evidence of cross-dataset generalization and potentially unfair comparisons. The paper does not evaluate cross-dataset transfer beyond the PartImageNet dataset family, despite emphasizing generalization to unseen parts. Cross-dataset evaluation is particularly important for establishing the transferability of open-vocabulary segmentation methods.

Response\color{#073}\textbf{Response}: We have added direct cross-dataset transfer for the evaluated PI-trained models, allowing us to distinguish foreground recovery from the transferability of part recognition.

Table 4 now evaluates PI-trained models directly on Pascal and ADE, without target training, using the complete target vocabularies.

PI-trained modelPascal All mIoUADE All mIoU
CLIPSeg†6.660.91
CAT-Seg†8.431.78
PartCLIPSeg†8.151.57
PartCATSeg†8.170.74
HOPS†10.780.66
PartFactor18.410.96

The contrast between Pascal and ADE is particularly informative. On Pascal, the correct part ranks first on 52.91% of valid annotated part pixels, including 12.96% subsequently rejected as background. On ADE, only \color{#B50}\textbf{4.22\\%} of pixels have the correct part ranked first, including 3.31% rejected by background (Appendix M.3).

This identifies a different bottleneck in ADE transfer. A foreground decision can retain a correct identity, but it cannot create the semantic correspondence that the decoder has not learned. The transfer experiment therefore helps distinguish where foreground learning is useful from where stronger recognition is needed.

AuthorsOpenReview ↗

Response to Reviewer ff9L (2/4)

W2.2. Contribution of stronger pretrained representations\color{#B50}\textbf{W2.2. Contribution of stronger pretrained representations}

Original comment:

Furthermore, PartFactor leverages multiple powerful pretrained feature extractors, including DINOv3-L, CLIP, SigLIP 2, and an ImageNet-supervised ResNet-50. Without controlling for differences in pretrained representations, it is difficult to determine whether the reported improvements arise from the proposed method or from the stronger combination of pretrained features.

Response\color{#073}\textbf{Response}: We examine pretrained representations separately from foreground adaptation, including a control using HOPS's pretrained initialization resources.

In Appendix Table 29, we replace the default encoders with HOPS's DINOv2-S/14 and CAT-Seg-initialized CLIP, removing DINOv3-L, SigLIP 2, and ResNet-50.

Under this initialization, the jointly trained PartFactor decoder obtains 39.44% unseen PI mIoU. Learning foreground on the same frozen decoder raises it to \color{#073}\textbf{41.29\\%, a gain of 1.85 points}.

This experiment matches the pretrained initialization resources while retaining PartFactor's architecture and frozen-encoder training policy. It demonstrates that the benefit of foreground specialization persists with a different representation.

The complete-system and fixed-decoder comparisons serve complementary purposes. The former measure overall segmentation performance, while the latter isolate foreground learning from changes to the part representation. We now distinguish these contributions explicitly throughout the manuscript.

Together, W2.1 and W2.2 establish that foreground recovery and transferable recognition are related but distinct problems. Understanding their respective contributions is one of the central goals of our analysis.

Weakness W3. Limited methodological novelty\color{#B50}\textbf{Weakness W3. Limited methodological novelty}

W3.1. Probabilistic factorization and background-logit reparameterization\color{#B50}\textbf{W3.1. Probabilistic factorization and background-logit reparameterization}

Original comment:

Limited methodological novelty. The central formulation decomposes the segmentation distribution into a foreground probability and a conditional distribution over part identities. The primary technical contribution is a background-logit reparameterization that makes foreground probability independent of part logits, conditioned on visual features. While mathematically sound, this is essentially a standard probabilistic factorization, and the reparameterization is equivalent to introducing an independent sigmoid foreground predictor.

Response\color{#073}\textbf{Response}: PartFactor uses a simple probabilistic construction to isolate a specific source of segmentation error and make its recovery experimentally measurable.

A missing part can arise from two different decisions. The decoder may rank an incorrect identity first, or it may rank the correct identity first and subsequently reject that pixel as background.

The probability factorization is standard, but it allows us to distinguish these errors precisely. Equation (3) shows why the highest-ranked part can still lose to background, while Equation (4) measures the event as correct-part rejection.

This distinction leads to a direct experiment: keep the conditional part distribution fixed on the decoder grid, learn only foreground support, and count the correct predictions that return.

Importantly, the failure persists after ordinary part–background supervision. On the jointly trained PI decoder, conditional unseen-part accuracy stays at 85.90%, while correct-part rejection falls from 18.94% to 17.10% after visual foreground adaptation (Appendix Table 27).

We also compare the two foreground decisions at matched background error. At a 2% background false-positive-rate budget, correct unseen-part recall rises from \color{#073}\textbf{62.19\\% to 70.95\\% on PI} and from \color{#073}\textbf{74.86\\% to 80.60\\% on OOD} (Section 5.3 and Appendix P). This measures an improvement in spatial selection beyond simply changing a global foreground threshold.

The mathematical construction is deliberately simple because it lets us intervene on the foreground decision without changing part recognition. This provides a direct way to investigate where useful predictions are lost.

AuthorsOpenReview ↗

Response to Reviewer ff9L (3/4)

W3.2. Advances beyond existing approaches\color{#B50}\textbf{W3.2. Advances beyond existing approaches}

Original comment:

The paper does not sufficiently establish how this formulation differs from, or advances beyond, existing approaches that separate class-agnostic localization from semantic recognition.

Response\color{#073}\textbf{Response}: Our contribution lies in connecting a measurable failure mode to a controlled recovery experiment, and establishing when a compact foreground intervention is effective.

To investigate the mechanism further, we introduce a controlled score-suppression experiment. Appendix Q supplies correct part identities, then subtracts the same spatial offset from all part logits. Their conditional ranking remains unchanged, but the original background decision becomes less reliable.

On PI, its unseen decoder-grid mIoU falls from 67.65% to 30.73%, while the visual foreground rule retains 67.48% under the same suppression (Table 37).

The experiment reveals how useful recognition can disappear through a later decision even when identity is correct. It also shows why the background reparameterization provides a clean way to study and correct this error.

The simplicity of PartFactor is deliberate. Once we can identify the decision responsible for losing a correct prediction, there is no need to relearn its identity merely to recover its spatial support. On PI, the foreground intervention requires only 22,177 trainable parameters\color{#073}\textbf{22,177 trainable parameters}, and its contribution can be measured without changing the decoder.

The important contribution is the connection between an otherwise obscured error, a controlled recovery experiment, and an economical way to improve the final prediction. The result also holds beyond a single trained decoder. Across three complete PI training runs with different decoder initializations, foreground adaptation improves unseen mIoU by 2.83, 5.59, and 3.58 points, averaging 4.00 ± 1.42 points\color{#073}\textbf{4.00 ± 1.42 points} (Appendix S, Table 42). Under matched caching, warm-up, and timing conditions, PartFactor also retains a 6.06× inference speedup\color{#073}\textbf{6.06× inference speedup} over HOPS (Table 41). These results connect the diagnosis to a repeatable, low-cost intervention and provide a concrete basis for deciding whether additional model capacity should target recognition or foreground selection.

Weakness W4. Insufficient discussion of related methodological precedents\color{#B50}\textbf{Weakness W4. Insufficient discussion of related methodological precedents}

Original comment:

Insufficient discussion of closely related methodological precedents. Although the related-work section covers several recent open-vocabulary part segmentation methods, it overlooks important prior work on decoupling class-agnostic segmentation and semantic classification. These studies are directly relevant to the proposed formulation and should be discussed to better contextualize the paper's novelty.

Response\color{#073}\textbf{Response}: We have expanded Section 2 to connect PartFactor directly to earlier work on grouping, recognition, and objectness.

ZegFormer, MaskCLIP, SAN, and FC-CLIP provide important precedents for separating spatial grouping from semantic recognition. OVRCOAT examines unseen-object suppression through objectness calibration, while DiSa studies foreground–background disentanglement.

These approaches establish the broader importance of treating recognition and spatial prediction as distinct components.

PartFactor investigates a specific consequence of their interaction in OVPS: which correctly ranked part identities are discarded by background, and how much segmentation can be recovered while those identities remain fixed?

The revised experiments connect this question to measurable rejection rates, matched-background-error curves, and foreground adaptation on an unchanged part decoder. They also show when the intervention is insufficient because the necessary identities are not available.

Section 2 now acknowledges the related methodological ideas while making the research question and experimental contribution of PartFactor clearer.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair: Summary of Revisions and Reviewer Discussion

Dear Area Chair,

We sincerely thank you and all three reviewers for the time, care, and thoughtful engagement devoted to our paper. We are especially grateful for your efforts in facilitating the discussion and encouraging reviewers to revisit the revised manuscript. Their feedback has helped us substantially improve both the research and its presentation.

We are encouraged that Reviewers ff9L and UVdM have raised their overall ratings from 2 (Weak Rejection) to 3 (Weak Acceptance) following the rebuttal. Reviewer ff9L indicated that our response resolved their concerns, while Reviewer UVdM found the response clear and considered most of their concerns addressed. We appreciate their willingness to reconsider the paper in light of new evidence. Their questions led to some of the most meaningful improvements in this work.

The revision is substantially more than a clarification of the original submission. We expanded the evaluation from two to four benchmarks, adding Pascal-Part-116 and ADE20K-Part-234, and conducted direct cross-dataset transfer experiments. We also redesigned the default training procedure to include ordinary part-and-background supervision from the beginning, directly testing whether the motivating failure persists under conventional training.

This experiment strengthened our understanding of the problem. Even after joint part-and-background training, a significant fraction of correctly identified unseen parts remains rejected as background. With the decoder frozen, foreground adaptation improves unseen mIoU across all four benchmarks. Three independently trained PI decoders show consistent gains, averaging 4.00 ± 1.42 percentage points (Table 42). Matched-background-error experiments further establish that the learned foreground decision improves the selection of correctly recognized parts, rather than merely shifting a global threshold.

We also added matched-pretrained-encoder experiments, architecture ablations, analyses of external decoders and small parts, additional foreground controls, and detailed evaluation-protocol comparisons. These experiments distinguish improvements arising from stronger recognition from those attributable to foreground adaptation. We have made the training configurations, evaluation code, trained models, and retained baseline checkpoints available for reproducibility.

Importantly, the scientific contribution has become clearer through these experiments. PartFactor begins with a question that is easy to overlook when evaluating only the final segmentation: when a part is missing, has the model failed to recognize it, or has a later decision discarded an identity it already knows? By formalizing correct-part rejection and investigating recovery with conditional part predictions fixed, we can identify where additional recognition learning is necessary and where a small, targeted foreground intervention is sufficient.

This understanding guides the design of the complete system. Under our unified evaluation, PartFactor achieves leading results against the retrained baselines on PI and OOD, with competitive performance on Pascal and ADE. It uses 9–21× fewer task-trainable parameters than the evaluated baselines, requires 4.5–13.1× less complete task-training time than HOPS, and achieves substantial measured inference speedups. Under matched caching and timing conditions, it remains 6.06× faster than HOPS on PI (Tables 5 and 41). These results demonstrate the practical value of directing learning toward a precisely identified source of error.

We appreciate Reviewer e7Up's acknowledgment that the rebuttal resolved part of their concerns, while recognizing that they remain unconvinced about methodological novelty. We responded with a further explanation of how PartFactor builds on existing work separating recognition, grouping, and objectness. Our contribution connects a measurable OVPS failure to controlled recovery experiments and an efficient complete system designed around that insight. We respect the reviewer's assessment and appreciate the opportunity to clarify this distinction.

Looking back, the review process has genuinely improved the paper. The additional experiments have helped us understand not only when PartFactor succeeds, but also why its benefits vary across decoders, which errors foreground learning cannot correct, and where stronger recognition remains necessary. These findings make the revised work more informative, better supported, and more useful to researchers studying open-vocabulary segmentation.

We are especially grateful to the reviewers for the scientific effort behind their critiques and to the Area Chair for guiding this discussion. We hope the final evaluation will consider the substantial changes to the manuscript, the new controlled evidence, and the reviewers' updated assessments.

Thank you again for your time, fairness, and thoughtful consideration of our work!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.