Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation

Thank you for taking the time to share your comments and suggestions in the review assessment.

We provide a detailed point-by-point response to your comments.

Q1: Contribution Clarity

A1:

Thank you for your insightful question. We would like to clarify the distinction between our two contributions:

• Our first contribution lies in proposing a new perspective for domain bridging, inspired by a theoretical insight that interprets image degradation as a mechanism for aligning domains. This conceptual contribution is general and does not rely on diffusion models, and it offers a novel way to understand how domain-invariant representations can emerge through degradation.
• Our second contribution provides a practical and effective instantiation of this perspective using the forward process of diffusion models. This implementation enables the construction of continuous intermediate domains, which progressively increase the domain overlap and facilitate robust adaptation.

In summary, the first contribution introduces a new theoretical lens for domain bridging, while the second demonstrates how to operationalize this concept effectively. The two are complementary yet distinct in novelty and impact.

Q2: Similarity to PiPa

A2:

We appreciate your reference to PiPa. As you correctly observed, both our method and PiPa adopt a student–teacher structure, which is a widely-used paradigm in self-training–based UDA methods, including DAFormer, HRDA, and MIC. This self-training framework is a general foundation, not a unique characteristic of PiPa.

However, our method DiDA totally differs from PiPa in both motivation, methodology, and learning objectives:

• Motivation: DiDA is driven by a theoretical insight grounded in degradation-based domain bridging, which is not explored in PiPa.
• Methodology: PiPa relies on pixel- and patch-wise contrastive learning, whereas DiDA constructs intermediate domains through image degradation and performs semantic shift compensation using a diffusion-based encoder.
• Learning Objectives: PiPa aims to improve representation through contrastive objectives, while DiDA focuses on learning from domain-shared distributions arising from degradation.

We will add a detailed comparison in the revised version to clarify these differences and avoid potential confusion.

Q3: Ambiguous Notations in Section 3.2

A3:

Thank you for pointing out the notational ambiguity. We would like to clarify that:

• The terms $g\_i$ and $g\_j$ appear exclusively in Section 3.2, which is dedicated to presenting our theoretical insight. These symbols represent modular attribute transformations used in the analysis of attribute granularity within the diffusion process. They are not part of the actual method design, nor do they appear in the method diagram (Figure 2). We will revise the text to make this distinction clearer.
• Regarding the term "attributes", as defined in Proposition 3.2, they refer to visual features of varying granularity, such as fine-grained textures or coarse-grained shapes. The proposition formally establishes that attributes with finer granularity are lost earlier in the diffusion process due to their higher sensitivity to noise. This analysis is theoretical in nature, and it provides a conceptual foundation for interpreting image degradation as a mechanism for domain bridging.

We appreciate this feedback and will revise Section 3.2 to better communicate the theoretical nature of this analysis and its role in motivating our design.

Q4: Section 3.2

A4:

We understand the confusion and thank you for the opportunity to clarify.

Section 3.2 is not a restatement of standard diffusion theory, but rather a presentation of new insight tailored to the UDA task. Specifically:

• We reinterpret the forward diffusion process from the perspective of attribute granularity, showing how different types of features (e.g., texture vs. shape) vanish at different time steps. This leads to the observation that the overlapping region of degraded distributions can serve as a domain-shared space.
• These insights are not part of the original diffusion model design, but are instead proposed as a novel conceptual framework for understanding how degradation can facilitate domain-invariant representation learning in UDA.
• This serves as a task-specific theoretical foundation and motivates the design of our DiDA framework, particularly the idea of constructing intermediate domains via degradation.

This task-specific reinterpretation of diffusion in the context of domain adaptation is, to our knowledge, novel and not previously explored. We will revise the section to make this reinterpretation more prominent.

Q5: On the Contribution of Degradation-Based Intermediate Domain Construction

A5:

Thank you for this question. While the mechanism is based on the forward diffusion process, the novelty lies not in the mechanism itself, but in its task-specific application to domain adaptation. That said, although the Degradation-based Intermediate Domain Construction module is indeed simple in implementation, it is not an isolated tweak, but rather a principled, task-specific strategy derived directly from the theoretical insight presented in Section 3.2. Specifically:

• We propose degradation not just as a noise process, but as a principled strategy to bridge domains by constructing a continuum of intermediate domains.
• The degradation levels are theoretically grounded (Section 3.2) and ensure that domain-specific features are gradually removed, while domain-invariant ones are preserved and compensated.
• The mechanism is analytically motivated, generalizable across datasets, and modular enough to be applied as a plug-in to existing UDA frameworks.

We believe this component offers a novel and insightful perspective on domain bridging in UDA. Importantly, while the implementation is concise, this simplicity reflects the elegance and universality of the underlying idea, rather than a lack of innovation.

Q6: Typos

A6:

Thank you for pointing this out. We will carefully review and correct all typographical errors in the revised version of the manuscript.

Q7: Missing SoTA Comparisons

A7:

We appreciate your suggestion. In Tables 3 and 4, we already compare DiDA against recent state-of-the-art UDA methods, including:

• ADFormer and CoPT, both of which are ECCV 2024 methods.
• Other recent methods such as MICDrop (also ECCV 2024) are also relevant, but our method already surpasses these works on the benchmarks we evaluate.

Additionally, as DiDA is designed as a plug-and-play module, we focused on integrating it into widely-used baselines such as DAFormer, HRDA, and MIC. Notably, MIC remains one of the strongest open-source UDA methods available, and DiDA consistently improves its performance.

We also note that DiDA, as a plug-in module, can be easily integrated into any future methods, which we plan to explore in follow-up work. In the final version, we will include comparisons with more recent methods, if they are released and relevant within the NeurIPS submission timeline.

Q8: Lack of Cross-City Evaluation

A8:

Thank you for this valuable suggestion.

We have extensively evaluated our method on multiple widely-used and challenging UDA benchmarks, including synthetic-to-real and clear-to-adverse-weather scenarios, which already demonstrate the robustness and generalizability of our approach.

While cross-city settings are indeed meaningful, we note that existing UDA methods have not been systematically evaluated under this setting, making it difficult to establish fair and informative comparisons. Moreover, the Cityscapes dataset itself already includes data collected from multiple cities, and its train/val split inherently introduces a cross-city domain gap. As such, the standalone “cross-city” setting may be relatively naïve in the context of UDA.

Nonetheless, we agree that explicit cross-city adaptation is a promising direction. We are currently extending DiDA to this setting and plan to include such results in future work.

We hope our response can resolve your concern. Please do not hesitate to let us know if you have further questions :)

Thank you.

1. Could you provide a formal theoretical demonstration for the first contribution? Otherwise, the first contribution is just an intuition for the second contribution.

2. Cross-City evaluation is adopted in multiple papers. So I still suggest the author to have a comparison if possible. [a] Hard-aware instance adaptive self-training for unsupervised cross-domain semantic segmentation. TPAMI 2025

Q2: Cross-City Evaluation

We apologize for the delay in our response. We greatly appreciate your valuable suggestion regarding the cross-city evaluation, and we have made our best effort to conduct the necessary experiments within the available time.

In response to your suggestion, we have followed the evaluation protocol of [a] “Hard-aware Instance Adaptive Self-Training” (TPAMI 2025), and conducted experiments on the Cityscapes → Oxford RobotCar benchmark, which exhibits a large domain gap in weather conditions. We include comparisons with both:

1. HIAST ([a]), which uses a DeepLabV2-based architecture, and
2. A more modern Transformer-based backbone (DAFormer), which demonstrates the applicability of our method to advanced UDA frameworks.

The results are summarized as follows (mIoU in %):

Observations:

• When applied to HIAST (DeepLabV2), our DiDA improves performance by +1.8% mIoU, showing its adaptability even with older architectures.
• When applied to DAFormer (MiT-B5), DiDA yields a +1.7% gain, achieving 81.4% mIoU, significantly outperforming prior methods in this setting.
• The improvements are consistent across all major object categories, including sign, light, and two-wheeler, which are known to vary across cities.

These results demonstrate that DiDA effectively bridges the domain gap in cross-city adaptation, and further confirms the generality of our approach across both conventional and modern architectures.

We will include this evaluation in the revised version and thank the reviewer again for the valuable suggestion.

We sincerely appreciate your thoughtful and constructive feedback throughout the review process. Your comments have helped us refine both the theoretical clarity and empirical completeness of our work. We believe the additional explanations and new experimental results have addressed your concerns, and we look forward to further improving the final version based on your insights.

Please feel free to let us know if any further clarification is needed.

Thank you.

Q1. So the first contribution is from [75]? How about your theoritical contribution? I still do not see a clear theory contribution ( proof or demonstration) to the problem.

Q2. Thank you. It is good to see that.

Considering Q1, I would like to keep my rating.

Dear author,

In your very-first reply, you claimed our theoretical insight multiple times.

So I asked you about the theoretical insight with formal proofment, definition and formulation.

Now you said it is not about proposing new theoretical theorems. But you have theoretical insight?

I think it is somehow confusing.

Thank you for taking the time to share your comments and suggestions in the review assessment. We appreciate your positive feedback on our novel Domain Bridging idea, principled module design, plug-and-play implementation, new state-of-the-art performance.

We provide a detailed point-by-point response to your comments.

Q1: Unverified Claim of Feature-Space Bridging

A1: Thank you for raising this insightful and important concern. The central hypothesis of DiDA is that progressive image degradation, inspired by the forward diffusion process, incrementally removes domain-specific features while preserving high-level, domain-invariant semantics. This process naturally constructs a sequence of intermediate domains that facilitate alignment between the source and target feature distributions.

=====================

First, we emphasize that our method is not only empirically motivated, but also grounded in a theoretical framework. Specifically, in Section 3.2, we provide a formal proposition (titled Attribute Loss and Time Step) that explains how different semantic attributes are progressively corrupted along the diffusion process. This proposition has two key implications:

• Fine-grained, domain-specific attributes (e.g., texture) are lost earlier in the diffusion process.
• Coarse-grained, domain-invariant attributes (e.g., shape) persist longer.

This supports our hypothesis that intermediate degradation levels (i.e., specific time steps) correspond to intermediate domain representations that preserve shared semantics while discarding domain-specific noise. As a result, samples from the source and target domains become more similar in feature space as degradation increases. Moreover, this theoretical insight justifies our core claim: image degradation acts as a transport mechanism that brings both domain distributions closer in latent space. It also motivates the diffusion encoder design, which compensates for semantic shift while preserving alignment.

=====================

To empirically validate this theoretical foundation, we provide a set of quantitative analyses in Appendix H (Fig. 7):

• In Fig. 7(a), we report model performance under both fully supervised and UDA settings across varying degradation levels. Models are trained independently in each distribution space, allowing us to isolate the effect of degradation on adaptation quality.
• Fig. 7(b) shows the relative UDA performance (i.e., UDA mIoU compared to the supervised upper bound) at each degradation level. As degradation increases, we observe a consistent rise in relative performance, supporting the idea that intermediate domains constructed by degradation better preserve domain-invariant structures and enhance domain alignment.

=====================

In addition, as illustrated in Section 4.3 and Fig. 4(b) of the main paper, we analyze inference behavior with fixed degradation levels. Two key findings emerge:

• Performance degrades gracefully with increasing diffusion steps, due to stronger corruption—this is expected.
• However, at each fixed degradation level, the model achieves its best performance when the inference step matches the training degradation level, and the resulting accuracy curves become flatter as degradation increases.

This suggests the model learns more robust, stable representations in deeper intermediate domains.

=====================

To provide more direct evidence of feature-level alignment, we have also conducted t-SNE visualizations of the encoder outputs:

• Before training, the source and target domains form clearly separated clusters.
• After DiDA training, these clusters become significantly more overlapped at low degradation levels (e.g., t = 0), and at higher degradation (t > 50), the source and target features nearly coincide.

This confirms that degradation serves as an effective bridge in feature space, not just pixel space. Due to the rebuttal format constraints, we are unable to include the t-SNE plots here, but we will incorporate them into the final version of the paper.

=====================

Together, these analyses—quantitative trends, behavioral patterns, and qualitative visualizations—strongly support that DiDA's gain stems not merely from regularization, but from a principled domain-bridging mechanism grounded in both theory and practice.

Q2: Class-wise Performance of Implicit Mode

A2:

Thank you for raising this thoughtful question. To better understand the source of the performance difference between implicit and explicit inference modes, we conducted a per-class mIoU comparison using a fixed diffusion step of $t = 50$ during inference. The results are summarized as follows:

We observe that implicit and explicit modes achieve comparable performance across most categories, with only minor fluctuations.

These results support our hypothesis that DiDA’s modeling is global and class-agnostic. There is no consistent or significant per-class bias introduced by either inference strategy. As described in Lines 309-326 and Figure 3–4 of the main paper, the two inference modes are functionally equivalent, differing only in how the semantic information is recovered at inference:

• The implicit mode uses a single forward pass through the degradation-aware network $\bar{f}_\theta$ to make predictions directly from noisy inputs and the known degradation level $t$.
• The explicit mode performs a two-stage process: it first predicts the added noise using $\hat{f}\_\theta$ reconstructs a clean image, and then passes it through the $f\_\theta$ for segmentation.

Despite this structural difference, both modes rely on the same underlying features and share weights, and both are trained with the same objective and data. Therefore, the observed performance gap is not due to optimization, but rather reflects the natural trade-off between robustness and feature disentanglement. Specifically:

• The explicit mode benefits from cleaner inputs after denoising, which slightly improves accuracy, especially for fine-grained classes.
• The implicit mode, on the other hand, demonstrates robustness to degradation and performs a more efficient, one-shot inference, which aligns with the training paradigm and reflects better generalization to noisy domains.

Additionally, both modes significantly outperform the base(S) and base(W) baselines, despite using a single shared model trained only once. This indicates that DiDA is capable of learning domain-invariant representations and performing semantic recovery implicitly, without needing explicit denoising at inference.

We will include these detailed per-class results in the appendix of the final version to provide more insight into the trade-offs between implicit and explicit inference. Thank you again for the thoughtful suggestion.

We hope our response can resolve your concern. Please do not hesitate to let us know if you have further questions :)

Thank you for taking the time to share your comments and suggestions in the review assessment. We sincerely appreciate your positive feedback on our rational and innovative approach, plug-and-play design, extensive experiments, and comprehensive ablation analysis.

We provide a detailed point-by-point response to your comments.

Q1: Increased training time due to the diffusion model

A1:

We fully acknowledge that incorporating a diffusion-based mechanism introduces additional training cost. However, we argue that this is a reasonable and justified trade-off for the following reasons:

• As detailed in Appendix E (Table 6), the training time increases from 14.5h to 18.5h on DAFormer and from 28.9h to 34.8h on MIC. This represents a moderate increase (~27%), especially considering the consistent performance gains of up to +3.7 mIoU across multiple settings.

• To reduce training cost, we propose a memory-efficient variant, DiDA($g_{time}$), which removes the diffusion encoder while preserving most of the performance (only 0.4 mIoU lower than full DiDA; see Table 2). This version reduces both memory consumption and training time significantly.

• Importantly, DiDA incurs no additional inference cost, as the diffusion-related modules are used only during training. This ensures deployment efficiency is unaffected.

• To further demonstrate DiDA’s efficiency, we compare its training cost and performance with other plug-in methods (e.g., FST, DTS) using the same DAFormer baseline and training schedule. Please refer to A4. Compared to FST and DTS, DiDA achieves better performance with lower training time and memory requirements, further validating its practicality.

Overall, we believe the performance–efficiency trade-off is well-balanced, and DiDA remains scalable and practical for real-world applications.

Q2: Lack of comparison with recent state-of-the-art methods

A2:

We appreciate your suggestion. In Tables 3 and 4, we already compare DiDA against recent state-of-the-art UDA methods, including:

• ADFormer and CoPT, both of which are ECCV 2024 methods.
• Other recent methods such as MICDrop (also ECCV 2024) are also relevant, but our method already surpasses these works on the benchmarks we evaluate.

Additionally, as DiDA is designed as a plug-and-play module, we focused on integrating it into widely-used baselines such as DAFormer, HRDA, and MIC. Notably, MIC remains one of the strongest open-source UDA methods available, and DiDA consistently improves its performance.

We also note that DiDA, as a plug-in module, can be easily integrated into any future methods, which we plan to explore in follow-up work. In the final version, we will include comparisons with more recent methods, if they are released and relevant within the NeurIPS submission timeline.

Q3: Limited improvement when combined with DTS

A3:

Thank you for raising this point. There appears to be a misunderstanding here.

We would like to clarify that in Tables 1, 3, and 4, DiDA is built directly on MIC, not DTS. DTS is listed for side-by-side comparison as another plug-in method.

Importantly, MIC is already a near-saturation UDA baseline, achieving 75.5 mIoU on GTA→CS. Our DiDA still manages to improve it to 76.8 mIoU (+1.3), which is non-trivial in this high-performance regime. As shown in prior works, gains >1 mIoU in such settings are considered meaningful and hard to achieve.

To avoid confusion, we will revise the description in the final version to make the experimental design and integration strategy clearer.

Q4: Training time comparison with other methods

A4:

As already shown in Appendix E (Table 6), we provide a detailed comparison of training time, GPU memory usage, and iteration cost. To offer further comparison, we provide a direct training cost comparison with other plug-in methods such as FST and DTS built on the same baseline (DAFormer) and training schedule (40K iterations):

This table highlights that DiDA not only achieves the best performance, but also maintains competitive training efficiency and lower memory usage than DTS and FST. Especially, DiDA($g_{time}$) offers a compelling balance between cost and accuracy, making it practical even under resource constraints.

We hope our response can resolve your concern. Please do not hesitate to let us know if you have further questions :)

Thank you for reminding me. After reviewing the other authors' responses I see that the author has provided very detailed answers and addressed most of the questions. Therefore I am raising my evaluation score to 5.

Dear Reviewer xKyR,

After reviewing the author's responses and the other reviews, could you share your updated thoughts following the rebuttal?

Best,

AC

I would like to thank the authors for carefully and thoroughly addressing my questions. This rebuttal has resolved most of my concerns. For these reasons, I would like to raise my evaluation score to 5. I wish you good luck.

Thank you for taking the time to provide detailed and constructive comments on our work. We greatly appreciate your positive feedback regarding the interesting and intuitive idea, well-motivated theoretical insights, good presentation, and solid and comprehensive experiments.

We provide a detailed point-by-point response to your comments.

Q1: Stability of The Framework

A1:

Thank you for highlighting this important concern. While DiDA introduces additional components to enhance domain adaptation, it has been carefully designed to ensure training stability, modularity, and practical efficiency. We address the concern from four perspectives:

1. Controlled Complexity of Loss Functions:
   DiDA introduces only two additional loss terms beyond standard self-training:

   • Degraded Image Consistency Loss ($L^D$), which regularizes predictions under degradation.
   • Reconstruction Loss ($L^R$), which supervises semantic recovery.

   Both are explicitly weighted in Eq. (10), and we conduct comprehensive ablation and sensitivity analysis in Appendix F (Tables 7 and 8), showing that DiDA maintains stable performance across a wide range of hyperparameter values. These losses are complementary and do not introduce optimization instability.

2. Modular and Loosely Coupled Components:
   As shown in Table 2, each module in DiDA can be independently removed or replaced without causing performance collapse. This modularity ensures robustness, and the architecture can be easily adapted or simplified depending on resource constraints.

3. Efficiency and Lightweight Variant:
   We provide a memory-efficient variant, DiDA ($g\_{time}$), which removes the diffusion encoder and retains only the time embedding. Despite its simpler design, it achieves comparable improvements (see Appendix E, Table 6), reducing GPU memory usage by up to 20%.

4. Inference-time Simplicity:
   Crucially, DiDA is used only during training. At inference, all auxiliary modules are disabled, and the model reverts to a standard segmentation network. This design ensures no additional latency, memory, or model complexity during deployment.

5. Empirical Generalization Across Setups:
   DiDA demonstrates consistent and significant improvements across diverse architectures (CNNs and Transformers), datasets (GTA→CS, SYN→CS, CS→ACDC), and UDA baselines (DAFormer, HRDA, MIC), as shown in Tables 1, 3, and 4. This cross-domain robustness supports the stability of our design.

While DiDA introduces two additional loss terms and auxiliary modules during training, the framework remains stable, modular, and efficient. The consistent performance gains across benchmarks validate its practical reliability.

Q2: Clarification of Intermediate Domain Construction

A2:

Thank you for pointing out the need for further clarification regarding intermediate domain construction.

The Degradation-based Intermediate Domain Construction module is based on the insight that gradually degraded images create a continuum of distributions between source and target domains. This strategy exploits the progressive removal of domain-specific features (e.g., texture) while retaining domain-invariant semantics (e.g., shape), thereby narrowing the domain gap in a controllable manner.

To clarify the mechanism:

• Figure 1 conceptually illustrates the forward diffusion process, where domain-specific information fades earlier (e.g., texture), and semantic shape information remains longer, under the guarantee of the theoretical insight (Section 3.2).
• Section 3.3 formally defines these intermediate domains as samples from the forward process $x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1 - \bar{\alpha}_t} \epsilon$, where $t$ controls degradation level.
• Figure 4 shows performance trends across different degradation levels, empirically validating the effectiveness of intermediate domains.
• Appendix H (Figure 7) presents a quantitative analysis: as degradation increases, the relative adaptation performance improves, suggesting that intermediate domains provide better alignment priors.
• Appendix L shows image-level reconstructions across noise steps (Figures 9–12), visually confirming that core semantics are preserved even under significant degradation.

We will revise the main paper to better explain the flow of degraded domain construction and directly reference these supporting sections.

Q3: Motivation for Semantic Shift Compensation

A3:

Thank you for raising this fundamental question. The Semantic Shift Compensation module addresses a critical limitation in degradation-driven adaptation: While degradation helps eliminate domain-specific cues, it may also corrupt domain-invariant semantics—especially under strong noise—leading to semantic shift and degraded predictions.

To mitigate this, we introduce a diffusion encoder ($g'$) and reconstruction head ($h'$), both conditioned on time-step embeddings. Together, they:

• Perceive the extent of degradation (via time embedding),
• Model the semantic shift caused by degradation,
• Recover lost semantic structure (supervised by reconstruction loss $L^R$),
• Regularize prediction consistency under noise (via $L^D$).

This design allows the model to preserve meaningful features even under strong degradation, as shown by:

• Appendix K (Figures 9–12): Visualizations of reconstructions under different degradation levels.
• Appendix M: Comparison with existing domain bridging methods shows our approach performs better in preserving semantics while reducing domain gap.
• Table 2 (Ablation): Removing this module leads to a noticeable drop in performance.

We will enhance the main text to better illustrate this motivation and its benefits.

Q4: Trade-off between Larger Timestep and Semantic Information

A4:

This is an excellent observation. Indeed, there is an intrinsic trade-off:

• Small $t$ values: low degradation → retain semantics but maintain domain gap.
• Large $t$ values: high degradation → reduce domain gap but risk semantic loss.

Rather than manually tuning this trade-off, DiDA leverages random sampling of $t$ during training (uniformly from $1$ to $T$), similar to standard diffusion training. More importantly:

• The time embedding module enables the model to perceive degradation level.
• The diffusion encoder ($g'$) adapts feature representations accordingly.
• The combined effect allows the model to implicitly learn how to balance domain invariance and semantic fidelity.

Empirical evidence:

• Figure 4(b) shows performance under fixed $t$: when $t$ is too large, semantic loss dominates, but the model performs best when $t$ matches degradation.
• Appendix H (Figure 7) shows relative performance gradually improves, which means that the model’s adaptability is strengthened across different $t$.
• Appendix K (Figure 10) confirms that the reconstruction is optimal when $t$ aligns with degradation.

This design ensures that the model learns to adapt across a spectrum of domain proximity and semantic fidelity, without requiring manual balancing.

Q5: Inference Process and Use of Compensatory Features

A5:

Thank you for pointing this out. We are happy to clarify the inference procedure and the usage of the diffusion encoder ($g'$) and reconstruction head ($h'$).

• Inference:

  • During inference, only the backbone segmentation network $f_\theta = h \circ g$ is used.
  • The diffusion-specific components—$g'$ and $h'$—are entirely removed at inference time.
  • This means no additional computation, parameters, or architecture changes are introduced in deployment, ensuring full compatibility with existing segmentation pipelines.

• Training:

  • The diffusion encoder $g'$ is used to model the semantic shift introduced by degradation by conditioning on the time step $t$. Its output is added to the main encoder’s output ($g$) at multiple feature levels via residual connections.
  • This residual fusion strategy avoids architectural disruption, helps preserve the original semantics, and makes the integration of $g'$ stable and lightweight.
  • This fused feature is then passed to:
    • the segmentation head $h$ for calculating the Degraded Image Consistency loss ($L^D$).
    • the reconstruction head $h'$ for the Reconstruction loss ($L^R$).

  This design ensures that:

  • The original encoder $g$ remains intact and continues to learn normally.
  • The auxiliary encoder $g'$ enhances the representation by compensating for semantic degradation in a plug-and-play fashion.
  • The training remains stable due to learnable residual modulation rather than full replacement or concatenation, which could otherwise destabilize feature distribution.

We will revise the main paper and Figure 2 to explicitly highlight this residual fusion mechanism and clarify that no concatenation or architectural modification occurs during inference.

Q6: Missing ACDC Results

A6:

Thank you for noticing this. The ACDC benchmark is indeed important as it evaluates domain shifts due to adverse weather.

• The main results for CS→ACDC are summarized in Table 1 (mIoU scores).
• Detailed per-class results are included in Appendix D, Table 5.

We agree that this benchmark deserves greater visibility. In the final version, we will move the key ACDC results and analysis into Section 4, ensuring clarity and completeness.

Q7: Typo

A7:

Thank you for pointing out this formatting issue. We will correct the extra parenthesis in Equation (7) in the final version.

We hope our response can resolve your concern. Please do not hesitate to let us know if you have further questions :)