acceptodds
Under review as a conference paper at ICLR 2027

GeoHaze: Geometry-Conditioned Spectral Propagation for Structure-Preserving Image Dehazing

Abstract

Image dehazing in severe natural haze and remote-sensing atmospheric degradation suffers from severe information suppression, making structure-consistent restoration depend on information beyond the degraded neighborhood. Monocular depth provides such nonlocal geometry, yet depth inferred from hazy inputs may contain shifted or missing boundaries; directly fusing it with RGB features can therefore entangle geometric errors with the appearance being reconstructed. We introduce GeoHaze based on a different principle: uncertain geometry should condition how RGB evidence propagates, rather than serve as reconstruction content. GeoHaze realizes this principle through a projected damped-wave spectral operator whose RGB state is driven by a spatially gated depth response, enabling global interaction while retaining RGB as the appearance carrier. Within this operator, two-level RGB–depth wavelet statistics predict sample-wise wave speed and damping, while a residual gate regulates how much of the propagated response re-enters the RGB stream; a frequency-decoupled decoder then uses geometric guidance to separately recover low-frequency layout and refine high-frequency detail, translating wave responses into structurally coherent reconstructions. Across 11 natural and remote-sensing benchmarks, GeoHaze achieves competitive restoration quality.Controlled wave variants and depth perturbations characterize the contribution and sensitivity of geometry conditioning, with an approximately 64% reduction in mean PSNR drop relative to SSDG across the three evaluated depth corruptions on RW²AH. Code is available in the supplementary materials.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image dehazing, real-world haze, remote sensing, vision foundation model, geometry prior, wave propagation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bTXeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GeoHaze, a geometry-conditioned spectral propagation framework for image dehazing. Building upon the damped-wave spectral operator introduced in WaveFormer, the proposed method incorporates monocular depth information through depth-conditioned initial velocity, wavelet-driven adaptive kinematics, and a geometry-gated frequency-decoupled decoder. Extensive experiments across natural and remote-sensing dehazing benchmarks demonstrate competitive performance.

Critical strengths
  1. The paper addresses a meaningful problem concerning the reliability of monocular depth priors in image restoration.

  2. The authors provide extensive architectural ablations, depth-corruption experiments, and implementation details.

Critical weaknesses

Weakeness 1

The authors emphasize the limitations of RGB-only image dehazing in the Introduction and Figure 1, suggesting that additional monocular depth guidance is important for structure-preserving restoration. However, this motivation is not sufficiently supported by the existing literature, as several RGB-only dehazing methods have already demonstrated substantially stronger performance on the same benchmarks.

For example, ConvIR [1] (TPAMI 2024) achieves PSNR values of 20.66 dB, 25.36 dB, and 22.44 dB on NH-HAZE, O-HAZE, and I-HAZE, respectively, outperforming GeoHaze (19.64 dB, 24.97 dB, and 19.14 dB) across all three datasets. Similarly, FocalNet [2] (ICCV 2023) achieves 20.43 dB on NH-HAZE and 25.50 dB on O-HAZE, while MB-TaylorFormer-v2 [3] (TPAMI 2025) reports 20.77 dB and 25.43 dB, respectively. Furthermore, Table 2 of MB-TaylorFormer-v2 includes several additional methods that achieve stronger results than those reported in Table 1 of this submission.

These observations raise two major concerns. First, the motivation regarding the inherent limitations of RGB-only dehazing appears overstated, since existing RGB-only approaches can achieve competitive or even superior restoration quality without explicit depth guidance. Second, the omission of these strong and well-established baselines makes the experimental comparisons incomplete and potentially misleading regarding the competitiveness of the proposed method.

While these results do not necessarily imply that depth guidance is ineffective, they substantially weaken the empirical justification for its necessity in the proposed framework. The authors should include comparisons with stronger RGB-only methods and provide more convincing evidence demonstrating the specific advantages of incorporating monocular depth guidance.

[1]Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. TPAMI, 2024.

[2]Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Focal network for image restoration. In ICCV, 2023.

[3]Jin, Zhi, et al. MB-TaylorFormer V2: Improved multi-branch linear transformer expanded by Taylor formula for image restoration. TPAMI, 2025.

Weakness 2

Equations (3)–(5) are overly concise and not sufficiently self-contained, with several mathematical notations left undefined or inadequately explained. Moreover, the formulation builds upon existing methods without clearly distinguishing the proposed contributions from prior work. As these equations describe the core methodology of GeoHaze, the insufficient explanations make it difficult to understand the technical details and identify the actual novelty of the proposed approach.

Questions for authors
  1. Clarification or experiments metioned from above Critical Weaknesses.
  2. Can the authors compare direct depth fusion and geometry-conditioned propagation within an identical architecture?
4 comments

Dear Reviewer bTXe,

Thank you for the constructive comments. We address your concerns below.


1. Strong RGB-only Baselines and Protocol Alignment

Concern: ConvIR, FocalNet, and MB-TaylorFormer V2 report higher PSNR than GeoHaze on NH/O/I-HAZE; omitting them makes the comparison incomplete and potentially misleading.

Direct answer. Thank you for the recommendation. We have added ConvIR, FocalNet, and MB-TaylorFormer V2 to the comparison and cited these works in the revised paper. We also identified and cited ArtIR (TIP 2026), whose efficient adaptive large-kernel design further supports our view that broad spatial context is important for image restoration. We disagree, however, with the claim that several RGB-only dehazing methods have already demonstrated substantially stronger performance on the same benchmarks, for three reasons: 1, PSNR alone does not comprehensively measure dehazing quality: structure, color cast, and perceptual detail matter, so we always report SSIM / LPIPS / MSE together with qualitative comparisons. 2, fair comparison requires a unified protocol; we retrain the listed RGB-only methods under our splits, crops, and tile inference next to GeoHaze, and the resulting scores can differ from the numbers reported in the original papers. 3, our goal is to show that introducing depth through the proposed gated-drive mechanism yields strong performance and cross-domain generalization on both natural-image and remote-sensing dehazing, whereas many existing methods remain strong mainly within a single domain.

Matched protocol. Published baselines and our setting differ in two ways: (1) train on all non-test images vs.\ our hold-outs (NH-HAZE 45/5/545/5/5; I-HAZE 25/5/525/5/5; O-HAZE 35/5/535/5/5); (2) our pipeline (384×384384\times384 crops, 512×512512\times512 tile inference) vs.\ release-specific full-resolution / resize. As a result, published ConvIR-B PSNR values you cite (22.4422.44 / 25.3625.36 / 20.6620.66 dB on I/O/NH-HAZE) should therefore not be line-compared to Table 1. Under this unified protocol we retrain ConvIR-B, FocalNet, and MB-TaylorFormer-v2-L from official code for the I/O-HAZE comparison; absolute scores differ from the papers and are kept separate from the released-checkpoint reevaluation in Appendix H.2. Every entry below shares the same experimental setting:

DatasetMethodVenueTypePSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowMSE ↓\downarrow
I-HAZEConvIR-BTPAMI'24RGB-only18.440.8198‾\underline{0.8198}0.3812‾\underline{0.3812}0.0151
I-HAZEFocalNetICCV'23RGB-only19.440.81600.38270.0141‾\underline{0.0141}
I-HAZEMB-TaylorFormer-v2-LTPAMI'25RGB-only18.500.81510.40900.0165
I-HAZEGeoHaze (Ours)—Depth-guided19.14‾\underline{19.14}0.82090.34070.0135
O-HAZEConvIR-BTPAMI'24RGB-only24.300.73430.33990.0083
O-HAZEFocalNetICCV'23RGB-only24.870.75500.36680.0074
O-HAZEMB-TaylorFormer-v2-LTPAMI'25RGB-only25.300.7861‾\underline{0.7861}0.37350.0040‾\underline{0.0040}
O-HAZEGeoHaze (Ours)—Depth-guided24.97‾\underline{24.97}0.78880.3658‾\underline{0.3658}0.0034

GeoHaze exceeds ConvIR-B on I/O-HAZE in PSNR (+0.70+0.70 / +0.67+0.67 dB) and is strongest or second on SSIM / LPIPS / MSE. On O-HAZE, MB-TaylorFormer-v2-L remains higher in PSNR. Relative to published ConvIR-B on I-HAZE (19.1419.14 vs 22.4422.44), the gap remains large and is not attributed entirely to protocol. The released-weight and matched-retraining evaluations are kept separate in Appendix H.2. Structure-oriented gains inside our backbone are in Sec. 2 (T3), not a claim that RGB-only dehazing is inherently weak.

Remote sensing. We have also investigated their performance on RRSHID-thin:

DatasetMethodPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowMSE ↓\downarrow
RRSHID-thinConvIR-B24.28950.68810.3648‾\underline{0.3648}0.005065
RRSHID-thinFocalNet24.4221‾\underline{24.4221}0.6884‾\underline{0.6884}0.36850.004776
RRSHID-thinMB-TaylorFormer-L-V224.05430.68210.38890.005379
RRSHID-thinGeoHaze24.430.69260.35390.0050‾\underline{0.0050}

Aa can be seen, GeoHaze leads this panel. Natural-haze PSNR advantages do not consistently transfer under this broader domain panel; geometry-conditioned GeoHaze stays strong on aerial RRSHID-thin.

Revision. We have included released-weight reevaluations in Tables 23–24 and matched-retraining results on I/O-HAZE and RRSHID-thin in Tables 25–26 in Appendix H.2 of the revised paper.

2. Motivation of Depth Guidance

Concern: Section 1 and Fig. 1 suggest that RGB-only dehazing is inherently limited, yet strong RGB-only methods already restore these benchmarks well. Please provide more convincing, specific advantages of monocular depth inside the proposed framework.

Direct answer. We would like to clarify the motivation again here. The question is conditional: monocular geometry is already used by depth-assisted dehazers. We take a step further to ask how imperfect geometry should enter the network.

Within-architecture ablation. All depth branches removed; same backbone as Full:

DatasetRGB-onlyFullΔ\DeltaPSNRΔ\DeltaSSIMΔ\DeltaLPIPSΔ\DeltaMSE
I-HAZE18.480019.1400+0.6600+0.6600+0.0189+0.0189−0.0298-0.0298−0.002216-0.002216
O-HAZE24.240024.9700+0.7300+0.7300+0.0218+0.0218−0.0292-0.0292−0.000622-0.000622
NH-HAZE18.940019.6400+0.7000+0.7000+0.0216+0.0216−0.0290-0.0290−0.002134-0.002134

Full improves all four metrics on each set. O/NH PSNR gains are only 0.070.07 / 0.040.04 dB above I-HAZE; we report an in-architecture depth-path benefit. On RW$^2$AH / RSID, shuffling depth costs U=1.1010U=1.1010 / 0.80000.8000 dB while Dˉ\bar D is only 0.06470.0647 / 0.05000.0500; Sc/US_c/U is lower than for SSDG. A clear-image DA3 depth proxy gains +0.2490+0.2490 dB on RW$^2$AH (diagnostic only). Depth is used at test time.

Revision. We have revised Sec. 1 and Fig. 1(a) to clarify the conditional role of imperfect geometry.


3. Self-Containedness of Eqs. (3)–(5) and Relation to WaveFormer

Concern: Equations (3)–(5) are overly concise and not sufficiently self-contained, with several mathematical notations left undefined or inadequately explained ... the insufficient explanations make it difficult to understand the technical details and identify the actual novelty of the proposed approach.

Direct answer. We agree that the submitted Eqs. (3)–(5) were too compact to read as a self-contained method description. In the revision, each equation is tied to an explicit role so that the technical story and the novelty boundary are both clear. Conceptually: Eq. (3) defines the 2D damped-wave equation on the feature grid, with wave speed vv and damping α\alpha; Eq. (6) gives the projected spectral update. The GeoHaze-specific insight is that imperfect monocular depth should not overwrite appearance as fused content, but condition that carrier. Concretely, RGB alone forms the propagated state U0U_0, while depth enters only as a gated drive V~0=Γd⊙ϕv(Fd)\widetilde V_0=\Gamma_d\odot\phi_v(F_d); sample-level (v,α)(v,\alpha) are predicted from joint RGB–depth Haar statistics (WDKE); and the residual path after the spectral step stays RGB-gated, with depth reused only as bounded frequency gates in decoding (GFDB). Thus the novelty of GeoHaze is geometry-conditioned driving and cross-modal kinematics on top of an existing spectral wave engine, with RGB remaining the appearance carrier. Every symbol is now defined at first use in Secs. 3.1–3.2, with Algorithm 1 (Appendix A.4) giving one HWPB forward pass and tensor shapes, and Table 10 / Appendix A.3–A.5 listing the full contract (Fr,FdF_r,F_d, ϕu,ϕv,ϕg,ϕo,ϕp\phi_u,\phi_v,\phi_g,\phi_o,\phi_p, sks_k, Δ\Delta, τ\tau, ϵ\epsilon, and the projected ρϵ\rho_\epsilon map).


4. Direct Depth Fusion vs. Geometry-Conditioned Propagation in One Architecture

Concern: Can the authors compare direct depth fusion and geometry-conditioned propagation within an identical architecture?

Direct answer. Yes. We replace only how depth enters the wave block, keeping backbone, WDKE/GFDB depth inputs, loss, and schedule fixed (shared panel with Reviewer rUxe).

Evidence.

DatasetVariantPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowMSE ↓\downarrow
RSIDC0 (depth as gated drive)27.09000.96250.02870.002700
RSIDC1 (RGB self-drive)26.50000.95400.04200.003093
RSIDC2 (depth as state content)26.63000.95600.03850.003002
RSIDC3 (concat replacement)26.55000.95480.04010.003057
RSIDC4 (cross-attention replacement)26.7200‾\underline{26.7200}0.9572‾\underline{0.9572}0.0362‾\underline{0.0362}0.002940‾\underline{0.002940}
RW$^2$AHC022.41000.66100.30110.010600
RW$^2$AHC121.93000.64100.32400.011836
RW$^2$AHC222.0400‾\underline{22.0400}0.6460‾\underline{0.6460}0.3192‾\underline{0.3192}0.011540‾\underline{0.011540}

C0 exceeds C2 by +0.4600+0.4600 / +0.3700+0.3700 dB (RSID / RW$^2$AH) and C4 by +0.3700+0.3700 dB on RSID. Zero-velocity is not the RGB-drive control. Full four-metric results are provided in Table 29 of Appendix H.3. Under these matched replacements, gated-drive is stronger than the listed content-fusion alternatives.


We appreciate your constructive commments agian and sincerely hope this clarification addresses your concerns.

Reviewer_bTXeOpenReview ↗

Response to rebuttal

I thank the authors for the detailed response and for taking the time to run additional experiments under a matched evaluation protocol. The rebuttal addresses most of my concerns, and the new comparisons with ConvIR, FocalNet, and MB-TaylorFormer-v2 show that GeoHaze is competitive on I-HAZE and O-HAZE. Given the modest PSNR gains over some baselines, I am still weighing whether the new results warrant raising my score. I will take the other reviewers’ assessments into account before finalizing my rating.

Dear Reviewer bTXe,

Thank you very much for your follow-up and for carefully reconsidering our submission. We sincerely appreciate your recognition that the matched comparisons address most of your concerns.

We acknowledge that the PSNR improvements on the natural haze benchmarks are modest and have revised the manuscript to avoid overstating performance. We would like to clarify once more here: Our main contribution is the geometry-conditioned interaction mechanism, which improves the same backbone over RGB-only and direct-fusion alternatives while remaining competitive across natural-image and remote-sensing settings.

For completeness, we have also corrected Appendix B.3 to clarify that GeoHaze and the newly evaluated baselines use the same tiled inference protocol at the original image resolution, without resizing or test-time augmentation.

We are grateful for your constructive comments, which have helped us improve the clarity and scope of the paper. If you have any remaining questions or concerns, we would be very happy to provide further clarification or address them.

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

This paper studies how to use error-prone monocular geometric priors to assist image dehazing. The authors point out that dense haze suppresses structural information in RGB, while depth estimated from hazy images may also contain boundary shifts or missing regions; directly fusing depth with RGB features can therefore carry geometric errors into the restoration process. To this end, the paper proposes to use geometric information to modulate the propagation of RGB features, rather than directly fusing depth as unconstrained appearance features.

The method builds on an existing projected damped-wave spectral operator and introduces a depth-driven propagation module (HWPB), a WDKE that predicts propagation parameters from RGB–depth wavelet statistics, and a geometry-gated frequency decoupling module (GFDB). The paper evaluates on 11 benchmark settings for natural-image and remote-sensing dehazing, and provides component ablations, propagation-parameter comparisons, depth perturbation tests, and downstream detection experiments.

Critical strengths
  1. The paper’s focus is not simply adding depth as an input, but how to leverage structural cues while limiting the influence of unreliable geometry. Depth driving, propagation-parameter modulation, and geometry-gated decoding together constitute the concrete implementation.

  2. The paper achieves competitive results across multiple natural-image and remote-sensing settings, and includes experiments that remove the depth encoder, remove the propagation module, replace it with local fusion, disable depth driving, fix the propagation parameters, and remove the geometric decoding gate. Depth perturbation tests further examine the model’s sensitivity to prior errors.

Critical weaknesses
  1. The existing local-fusion replacement provides relevant support for geometry-conditioned propagation, and the authors explicitly acknowledge that their results do not establish superiority over every attention-fusion method. Nevertheless, differences among the external methods in backbone, module structure, and training configuration, together with the limited range of internal fusion alternatives, make it difficult to fully distinguish whether the advantage comes from the way geometric information is used, the global spectral operator, or their combination with the decoding structure. A stronger matched fusion alternative could further strengthen this attribution. Since depth enters the feature update through the driving term and gating, a more explicit explanation of the structural constraints distinguishing this interaction from direct fusion would also help connect the design principle to the implementation.

  2. The paper has explicitly limited its robustness conclusions to the three perturbations tested. However, the lower sensitivity to perturbations could stem from the propagation structure, the gating, the depth augmentation during training, or differences in reliance on depth information. The existing depth ablations support the usefulness of depth, but the cross-method comparison of performance drops does not fully distinguish the benefit obtained from useful depth information from sensitivity to corrupted depth. This limits attribution of the robustness improvement to the propagation mechanism itself.

Questions for authors
  1. The paper introduces depth perturbation or augmentation during training. Did the SSDG used in the robustness comparison receive the same type and strength of training augmentation? If not, should the current results be understood as a comparison between complete training schemes rather than an isolated advantage of the propagation mechanism?

  2. Which information paths are specifically removed by “removing the depth encoder”? In Table 6, this variant still retains WDKE. Please explain how WDKE’s input is constructed in this case, and how depth driving, geometric gating, and any depth-derived statistics are handled. In particular, should this variant be interpreted as removing the depth encoder specifically, or as removing all geometric information?

  3. The experiments that fix the propagation parameters and disable depth driving already provide component-level evidence. Could additional analysis clarify what incremental information the RGB and depth statistics provide to WDKE? For example, matched input-source comparisons or an analysis of the predicted parameters in relation to depth quality and haze intensity could help explain their respective roles.

4 comments

Dear Reviewer rUxe,

Thank you for focusing on how to use structural cues while limiting unreliable geometry, for the depth-perturbation and ablation coverage you noted, and for treating our claims as appropriately scoped. We address your concerns below.


1. Did SSDG Receive the Same Depth Augmentation?

Concern: Did SSDG in the robustness comparison receive the same type and strength of training augmentation? If not, should Table 4 be read as a comparison of complete training schemes rather than an isolated advantage of the propagation mechanism?

Direct answer. No. In the submitted manuscript, SSDG was retrained with its official recipe and Depth Anything 3 priors, without our depth-perturbation augmentation. Table 4 in the manuscript therefore compares complete training schemes, as you suggest. We agree that this comparison alone does not isolate the propagation mechanism.

Evidence. On a same-backbone 2×2×22\times2\times2 factorial of global vs. windowed spectral mixing (PP), HWPB geometric gating (GG), and depth augmentation (AA) on RW$^2$AH, average main-effect Δ\DeltaPSNR is P=+0.5753P=+0.5753, G=+0.2203G=+0.2203, A=+0.1153A=+0.1153. The unaugmented full operator (110: 22.279022.2790 dB) still exceeds the augmented windowed/ungated setting (001: 21.599021.5990 dB). With matched training-time depth augmentation switched on/off within each method, mean absolute test-time corruption sensitivity Dˉ\bar D drops from 0.12000.1200 to 0.06470.0647 for GeoHaze and from 0.17950.1795 to 0.10510.1051 for SSDG: augmentation helps both schemes but does not by itself equalize them. The full eight-row factorial is in the revised appendix.

Revision. We have revised Sec. 4.5, which states that the original Table 4 compares complete schemes more clearly. We have included the within-method training-time depth-augmentation comparison in Table 34 of Appendix H.8 of the revised paper.


2. What Separates Geometry-Conditioned Interaction from Direct Fusion?

Concern: External methods mix backbone, modules, and training, so it is hard to tell whether gains come from how geometry is used, from the global spectral operator, or from their combination with the decoder. Stronger matched fusion alternatives and an explicit account of the structural constraints that distinguish this interaction from direct fusion are needed.

Direct answer. Thanks for the constructive comment and we would like to clarify the advantage of our proposed mechanism here. The mechanism advantage is how geometry is used, not that we add a stronger backbone or decoder. Depth does not enter as fused appearance content (concat / cross-attention style). It enters as a gated drive of a global spectral filter: U0U_0 stays RGB-only, depth supplies V~0=Γd⊙ϕv(Fd)\widetilde V_0=\Gamma_d\odot\phi_v(F_d) plus sample-level (v,α)(v,\alpha), and the operator applies distinct responses H0H_0 (RGB) and H1H_1 (depth drive), with ∣H1∣|H_1| decaying as ∼1/(vκ)\sim 1/(v\kappa) at high frequency. That H0/H1H_0/H_1 split is the structural constraint that direct fusion does not have. Under a matched protocol that freezes backbone, WDKE, GFDB, loss, and schedule, and changes only the depth slot in HWPB, gated-drive C0 beats RGB self-drive (C1), depth-as-content (C2), and same-slot concat / cross-attention (C3/C4) by +0.37+0.37–0.590.59 dB PSNR. The same panel also separates the three confounded sources you raise: the global operator helps (PP main effect +0.5753+0.5753 dB), yet C1 already has that operator and still trails C0, so how geometry is used matters beyond the operator alone; GFDB is shared across C0–C4, so the gaps are not from a decoder unique to C0.

Structural constraint. Appendix Eq. (14), together with main-paper Eq. (7), gives, for radial frequency κ\kappa with ρ≈vκ\rho\approx v\kappa,

|H_0(\kappa)|\le e^{-\alpha\tau/2}\Bigl(1+\tfrac{\alpha}{2}\min\bigl(\tau,1/\rho\bigr)\Bigr).$$ RGB is filtered by $H_0$; the depth is driven by $H_1$. Concat / cross-attention fusion has no such split. **Matched panel (PSNR).** C0: gated depth drive. C1: same global operator, RGB self-drive (not RGB-only). C2: depth as state content. C3/C4: concat / cross-attention in the same slot. | Dataset | C0 | C1 | C2 | C3 | C4 | | --- | ---: | ---: | ---: | ---: | ---: | | RSID | 27.0900 | 26.5000 | 26.6300 | 26.5500 | 26.7200 | | RW\$^2\$AH | 22.4100 | 21.9300 | 22.0400 | — | — | Equal-RMS band-limited depth noise further aligns with the mechanism: C0 high-band drop is $0.0210$ dB vs.\ $0.1800$ / $0.2000$ for C1/C2 ($\sim|H_1|$). **Revision.** We have added the $H_0/H_1$ decomposition and bounds after Eq. (6) in Sec. 3.2 (Eq. (7)). The full C0–C4 results are provided in Tables 28–29 of Appendix H.3 of the revised paper and summarized in Sec. 4.4. We have included the equal-RMS high-band depth-noise results in Table 33 of Appendix H.7 of the revised paper. ---

3. Source of the Lower Depth Sensitivity

Concern: Lower sensitivity could come from propagation structure, gating, training augmentation, or lower reliance on depth. Cross-method Δ\DeltaPSNR drops do not fully separate useful-depth benefit from corruption sensitivity.

Direct answer. The lower sensitivity is not from using less depth. We separate useful-depth benefit from corruption sensitivity with two depth-only interventions on frozen weights: U=PSNR(clean)−PSNR(shuffle)U=\mathrm{PSNR}(\mathrm{clean})-\mathrm{PSNR}(\mathrm{shuffle}) (and UcU_c for constant-map depth) measures how much aligned geometry helps, while ScS_c and Dˉ\bar{D} measure drops under the three tested corruptions. On RW$^{2}AH,C0hasalargeruseful−depthbenefitthantheSSDGanchor(AH, C0 has a larger useful-depth benefit than the SSDG anchor (U=1.1010$ vs. 0.94700.9470) and a much smaller corruption sensitivity (Dˉ=0.0647\bar{D}=0.0647 vs. 0.17950.1795, about 63.94%63.94\% lower), with every Sc/US_c/U ratio also lower. A horizontal depth flip still costs 0.85100.8510 dB, so the model depends on scene-aligned geometry rather than ignoring it. Among the remaining candidate sources, training augmentation helps but is not dominant (factorial main effects A=+0.1153A=+0.1153 vs.\ P=+0.5753P=+0.5753, G=+0.2203G=+0.2203; Sec.~1), and appendix C1/C2 rows under the same probe remain more corruption-sensitive than C0. We have conducted complementary evaluation as follows,

Evidence. Probe on RW$^2$AH:

ModelUUUcU_cDˉ\bar{D}S(noise)/US(\mathrm{noise})/US(blur)/US(\mathrm{blur})/US(edge)/US(\mathrm{edge})/U
C0 (GeoHaze)1.10101.25100.06470.02360.07110.0817
SSDG (anchor)0.94701.09700.17950.34140.10720.1201

A clear-image DA3 depth proxy gains +0.2490+0.2490 dB (diagnostic only). RSID replication and C1/C2 rows are in the revised appendix.

Revision. We have revised Sec. 4.5 and reported the depth-reliance/sensitivity probe in Table 32 of Appendix H.6 of the revised paper. The corruption operators are defined in Appendix E.1. Robustness wording stays limited to the tested perturbations.


4. What Does “w/o Depth Encoder” Remove?

Concern: Which paths does “removing the depth encoder” cut? Table 6 still keeps WDKE. Is this ablating the encoder only, or all geometric information?

Direct answer. It ablates multi-scale depth feature learning, not every geometric pathway. It should not be read as a full removal of geometric information.

Evidence. With the depth encoder disabled: patch embedding / encoder / downsampling are bypassed and replaced by zero tensors at each scale; WDKE still runs with normalized zero placeholders on the depth branch; depth-drive and decoder projections remain instantiated; the HWPB depth gate is bypassed. Table 6 reports 26.370026.3700 dB vs.\ Full 27.090027.0900 dB on RSID. As a full-removal reference, RGB-only GeoHaze (all depth branches stripped) on I/O/NH-HAZE trails Full by +0.6600+0.6600 / +0.7300+0.7300 / +0.7000+0.7000 dB PSNR.

Revision. We have renamed the row in Table 6 to “w/o depth feature encoder” and clarified the remaining depth paths in Sec. 4.5. The full-removal RGB-only reference is provided in Table 27 of Appendix H.2 of the revised paper.


5. Incremental Roles of RGB and Depth Statistics in WDKE

Concern: Please clarify what incremental information RGB vs.\ depth statistics provide.

Direct answer. Under matched training on RW$^2$AH, joint RGB–depth Haar statistics outperform RGB-only, depth-only, spatial GAP, and fixed (v,α)(v,\alpha) by +0.2900+0.2900 / +0.3600+0.3600 / +0.4500+0.4500 / +0.2300+0.2300 dB PSNR. Predicted (v,α)(v,\alpha) vary across blocks within the design bounds; because vv sets the high-frequency bandwidth of the depth drive through H1H_1, this is consistent with sample-wise bandwidth calibration, not with depth-error detection. The full input-source table is in the revised appendix next to the fixed-(v,α)(v,\alpha) ablation.

Revision. We have added the WDKE input-source comparison in Table 31 of Appendix H.5 of the revised paper and summarized its results in Sec. 4.5.


We appreciate that these comments helped us separate augmentation from architecture, isolate geometry-as-drive under matched fusion controls, make the H0/H1H_0/H_1 structural constraint explicit, and clarify depth reliance versus corruption sensitivity together with the WDKE input roles. We sincerely hope this clarification addresses your concerns.

Thank you for the additional experiments. I noticed an inconsistency in the evaluation protocol: Appendix B.3 of the original submission explicitly states that inference uses neither tiling nor resizing, while your response to bTXe lists 512×512 tiled inference as part of the unified comparison protocol.

Please clarify the inference protocol used for GeoHaze and each baseline in the original Table 1 and the newly added baseline comparison. Was GeoHaze reevaluated under the same protocol as the new baselines? Its I-HAZE and O-HAZE scores remain 19.14 and 24.97, respectively. Please specify the evaluation settings corresponding to these values and clarify whether the original description needs correction or the new experiments use a different protocol.

Dear Reviewer rUxe,

Thank you for identifying this inconsistency. We carefully checked the relevant statements and re-examined our experimental logs, and found that Appendix B.3 of the original submission incorrectly stated that the inference used no tiling. The statements concerning no resizing and no test-time augmentation were correct. We overlooked this sentence when preparing the revision and apologize for the confusion.

For both the original GeoHaze results and the newly added matched comparison, all images retain their original resolution and all four methods use the same FP32 evaluation protocol: 512×512512\times512 tiles, 64-pixel overlap, normalized Hann blending, no resizing, and no test-time augmentation. We adopted this common protocol because full-image inference under the same 24 GB memory limit produced OOM in some images of the O-Haze and I-Haze datasets (i.e., images in these two datasets are at different resolutions) for ConvIR-B, FocalNet, MB-TaylorFormer-L V2, and GeoHaze, respectively. Thus, applying tiling to every method avoids mixing inference settings. Therefore, the GeoHaze scores of 19.14 on I-HAZE and 24.97 on O-HAZE already correspond to this tiled protocol and remain unchanged in the new comparison.

The revised manuscript and our response to Reviwer bTXe now state this protocol explicitly. We will also correct Appendix B.3. This was an error in documenting the original setting, not a change in the protocol used to obtain the reported results.

We would like to express our sincere gratitude once again for the care and diligence you demonstrated throughout the review process.

Reviewer_7CicOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes GeoHaze for single‑image dehazing for both natural hazy scenes and remote‑sensing haze removal. Different from existing depth‑assisted dehazing methods that directly fuse depth maps as visual features into RGB streams, this work advocates a core principle: imperfect depth priors should condition feature propagation rather than act as reconstruction content. Built upon WaveFormer’s damped wave spectral operator, the authors design a Heterogeneous Wave Propagation Block (HWPB), a Wavelet‑Driven Kinematics Estimator (WDKE), and a Geometry‑Gated Frequency‑Decoupling Block (GFDB). Depth features spatially gate the initial driving velocity of the wave PDE, and sample‑wise wave speed / damping coefficients are predicted from joint RGB‑depth wavelet statistics. The frequency‑decoupled decoder separates low‑frequency layout recovery and high‑frequency detail refinement. Extensive experiments are conducted across 11 benchmarks including real‑world indoor/outdoor haze and multiple remote‑sensing cloud‑haze datasets.

Critical strengths

The observation that direct depth‑RGB feature fusion propagates depth estimation artifacts into restored outputs is practically meaningful for real‑world dehazing, since monocular depth predicted from heavily hazy inputs inevitably contains boundary shift and missing structures. Instead of discarding depth cues entirely, the authors propose to use depth as a propagation controller rather than appearance content, which offers a promising alternative paradigm for geometry‑prior‑based restoration.

Critical weaknesses
  1. The paper reports parameter count (11.49 M) and FLOPs (57.64 G), but does not report actual wall‑clock inference latency, memory footprint, GPU peak memory. Spectral‑domain operations (2D‑FFT / iFFT inside every HWPB block) bring non‑trivial real‑world overhead, especially for high‑resolution 4K images. For airborne / remote‑sensing scenarios targeted in this paper, latency and memory are critical constraints. Readers cannot judge whether the FFT‑based spectral operator is deployable for real‑time applications. Please add inference speed and peak memory comparisons against representative baselines (DehazeFormer, PoolNet, SSDG, WaveFormer) on same hardware.

  2. The authors adopt damped wave‑equation as a feature evolution tool. However, they explicitly state that it is a projected spectral filter instead of strict physical wave propagation on scene geometry. Some wording in main text may mislead readers into believing features propagate along real‑world geometric surfaces. The authors should carefully revise wording throughout Section 3 to emphasize this is a PDE‑inspired spectral filter for feature mixing, not physical wave travelling inside 3D scene space.

  3. Limitation discussion is too brief. Expand failure‑case analysis: uniform dense haze, repetitive thin textures, monocular depth ambiguous scenes, as mentioned in the text. More concrete qualitative failure examples in appendix are preferred.

Questions for authors

The authors adopt damped wave‑equation as a feature evolution tool. However, they explicitly state that it is a projected spectral filter instead of strict physical wave propagation on scene geometry. Some wording in main text may mislead readers into believing features propagate along real‑world geometric surfaces. The authors should carefully revise wording throughout Section 3 to emphasize this is a PDE‑inspired spectral filter for feature mixing, not physical wave travelling inside 3D scene space.

2 comments

Dear Reviewer 7Cic,

Thank you for recognizing geometry as a propagation control rather than appearance content as a promising alternative paradigm, and for the clear, actionable requests on efficiency, Section 3 wording, and failure cases. We address your concerns below.


1. Does Section 3 Describe 3D Surface-Wave Travel, or Only 2D Feature-Grid Filtering?

Concern: The wording in Section 3 can be read as physical wave travel along 3D scene surfaces. Please clarify the scope of “propagation” and make the operator’s domain precise.

Direct answer. We agree that parts of Section 3 could suggest waves travelling on 3D surfaces and appreciate that you point out this issue. The operator is a projected spectral filter instead of strict physical wave propagation on scene geometry, as you mentioned.

Revision. We have clarified these claims in the revised manuscript. Sec. 3.1 defines U(x,τ)U(x,\tau) on the flat 2D feature grid and separates the depth-gated drive from WaveFormer’s homogeneous RGB carrier. Sec. 3.2 describes the operator as a “PDE-inspired projected spectral filter in the 2D Fourier domain.” Fig. 2 identifies HWPB as a “PDE-inspired projected spectral filter.” Fig. 3’s caption places trajectories in PCA feature space. Appendix A.4 uses “depth-gated initial drive” rather than 3D surface motion.


2. Is the FFT-Based Operator Practical at High Resolution, Including 4K?

Concern: Repeated 2D-FFT/iFFT inside HWPB may incur non-trivial overhead at high resolution. Please report real-world latency and memory against DehazeFormer, PoolNet, SSDG, and WaveFormer, including 4K, and clarify FFT cost.

Direct answer. We measured wall-clock latency and peak memory on one GPU. HWPB runs only at interaction stages with counts [1,1,2][1,1,2] (Appendix Table 11), i.e., after spatial downsampling. At full 4K without tiling, the median per-run 2D-FFT/iFFT share of the restoration network’s GPU (NVIDIA A30) active time is 2.07%; end-to-end cost is dominated by the rest of the network and, when included, by Depth Anything 3.

Table A. 4K inference with AMP (2160×38402160\times 3840). Latency in ms; memory in GiB.

ModelSpeed ↓\downarrowGPU alloc. ↓\downarrowGPU reserv. ↓\downarrowCPU RSS ↓\downarrow
SSDG + Depth Est.OOM———
PoolNet (fixed)3318.1717.14319.9361.097
DehazeFormer3207.6613.33121.4201.466
WaveFormer2981.6817.46221.9471.396
Ours + DA32050.4510.32113.2322.496
Ours (w/o DA3)1780.739.07511.7711.404

As shown in this table, under 4K AMP, Ours+DA3 reduces median latency by 36.1% (1.56$\times$) and peak allocated memory by 22.6% relative to DehazeFormer. For larger remote sensing scenes, overlapping tiles remain a standard memory option; because HWPB already runs on downsampled feature maps, tiling is not required to hide FFT cost.

Revision. We have included the 4K AMP results in Table 8 of the main text and Table 22 in Appendix H.1 of the revised paper. The full resolution sweep and operator profile are also provided in Appendix H.1. Table 20 reports median component times and median per-run shares separately.


3. Can You Show Concrete Failures for the Modes Named in the Paper?

Concern: The limitation discussion is too brief. Please expand failure-case analysis for uniform dense haze, repetitive thin textures, and monocular-depth-ambiguous scenes, preferably with qualitative examples in the appendix.

Direct answer. We agree and have expanded Sec. 4.6 and Appendix I with concrete qualitative examples of the limitations named in the text.

Evidence. Sec. 4.6 now states the boundary, where GeoHaze can fail when RGB and depth cues are uninformative, including large-area depth hallucination, near-uniform dense haze, thin repetitive textures, and ambiguous monocular depth in near-orthographic scenes (As can be seen in Appendix I, Fig. 14). Main Fig. 7 shows full images with enlarged local insets, while Appendix Fig. 14 retains the full comparison. Three RW$^2$AH candidates with matched crops (Input, GT, Full, Cross attention): (a) uniform dense haze 1-0097-01; (b) monocular-depth ambiguity 1-0199-05; (c) repetitive fine textures 1-0083-01. We also show reflective water 12-0001-06 and an auxiliary sky-boundary example 8-0026-01. Complementary quantitative depth-corruption results remain in Table 4 and Appendix E.

Revision. Large-area depth hallucination (sky / water) is already listed in Sec. 4.6. Fig. 15 has been added to Appendix I, with a corresponding reference in Sec. 4.6.


We appreciate your constructive comments and sincerely hope this clarification address your concerns.

Dear Reviewer 7Cic,

Thank you again for your time and constructive feedback.

With about several hours remaining in the author–reviewer discussion period, we would greatly appreciate it if you could take a moment to review our rebuttal and revised manuscript. We have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and welcome your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

7 comments
AuthorsOpenReview ↗

Summary of Revisions in the Updated Manuscript

We thank the reviewers for their constructive suggestions. The revised manuscript incorporates the changes below:

  1. Operator-domain wording (Reviewer 7Cic).
    Sections 3.1–3.2 and Figure 2 now describe HWPB as a PDE-inspired projected spectral filter on the flat 2D feature grid / 2D Fourier domain, rather than physical wave travel on 3D scene surfaces. Figure 3’s caption places trajectories in PCA feature space.

  2. Self-contained equations and novelty boundary (Reviewer bTXe).
    Symbols in Secs. 3.1–3.2 are defined at first use; Eq. (6) gives the projected spectral update and Eq. (7) the H0/H1H_0/H_1 envelopes. Table 9 separates inheritance from WaveFormer from GeoHaze-specific additions. Appendix A.3–A.5 and Algorithm 1 record the HWPB contract and one forward pass.

  3. Matched depth-role controls C0–C4 (Reviewers rUxe, bTXe).
    Section 4.4 and Appendix H.3 (Tables 28–29) compare gated depth drive, RGB self-drive, depth-as-content, concat, and cross-attention under a frozen backbone / WDKE / GFDB / loss / schedule protocol. Appendix H.7 (Table 33) reports equal-RMS high-band depth-noise drops.

  4. Robustness attribution and depth ablations (Reviewer rUxe).
    Section 4.5 clarifies that Table 4 compares complete training schemes. Appendix H.4 / H.8 (Tables 30, 34) report the PP-GG-AA factorial and training-augmentation toggle; Appendix H.6 (Table 32) separates useful-depth benefit from corruption sensitivity. Table 6 is renamed to “w/o depth feature encoder,” with full RGB-only GeoHaze in Appendix Table 27.

  5. WDKE input sources (Reviewer rUxe).
    Appendix H.5 (Table 31) compares joint RGB–depth Haar statistics with RGB-only, depth-only, spatial GAP, and fixed (v,α)(v,\alpha) under matched training.

  6. RGB-only baselines and claim scoping (Reviewer bTXe).
    Section 1 and Figure 1(a) now pose a conditional question about imperfect geometry rather than claiming that RGB-only dehazing is inherently limited. Appendix H.2 reports released-weight reevaluation (Tables 23–24), matched retraining on I/O-HAZE and RRSHID-thin (Tables 25–26), and within-architecture RGB-only GeoHaze (Table 27). ArtIR is cited as related large-kernel restoration context.

  7. Inference-protocol documentation (Reviewer rUxe).
    Appendix B.3 corrects the original overstatement: I/O/NH-HAZE evaluations in Table 1 used FP32 native-resolution inference with 512×512512\times512 tiles, 64-pixel overlap, and normalized Hann blending, without resizing or test-time augmentation. The matched I/O-HAZE panel in Table 25 uses the same tiled protocol; RW$^2$AH remains full-image.

  8. High-resolution efficiency (Reviewer 7Cic).
    Table 8 and Appendix H.1 (Tables 20–22) report 4K AMP latency / memory against DehazeFormer, PoolNet, SSDG, and WaveFormer, together with FFT/operator profiling.

  9. Failure cases (Reviewer 7Cic).
    Section 4.6 expands the limitation discussion, and Appendix I / Figure 14 provide qualitative examples for uniform dense haze, monocular-depth ambiguity, and repetitive fine textures.

We hope these revisions address the concerns and look forward to the discussion.

Authors of Submission 1433

AuthorsOpenReview ↗

General Response & Rebuttal Summary

Dear Area Chairs,

To make the discussion easier to navigate, we first summarize what the reviewers highlighted as strengths, then group the shared concerns and point to where each is handled in the individual rebuttals and in the revised manuscript.

  • R1: Reviewer 7Cic
  • R2: Reviewer rUxe
  • R3: Reviewer bTXe

Key strengths:

  1. The core principle is meaningful for real-world dehazing: imperfect monocular depth should condition feature propagation rather than act as unconstrained appearance content (R1, R2, R3).
  2. Depth driving, propagation-parameter modulation, and geometry-gated decoding provide a concrete implementation of that principle, with competitive results across natural-image and remote-sensing settings (R2, R3).
  3. The evaluation includes component ablations, depth-perturbation tests, and implementation detail that support controlled reading of geometry conditioning (R2, R3).

Key concerns:

  1. Method wording, operator domain, and novelty boundary (R1 W2/Q, R3 W2).
    Reviewers asked whether Section 3 describes physical 3D surface-wave travel and whether Eqs. (3)–(5) are self-contained enough to identify what GeoHaze adds beyond WaveFormer. In the revision, Sec. 3.1 defines U(x,τ)U(x,\tau) on the flat 2D feature grid, Sec. 3.2 / Fig. 2 describe HWPB as a PDE-inspired projected spectral filter in the 2D Fourier domain, and Eq. (7) states the H0/H1H_0/H_1 envelopes that separate RGB-state filtering from the depth drive. Table 9 and Secs. 3.1–3.2 further separate inheritance from WaveFormer from the geometry-conditioned additions (gated drive, WDKE, GFDB). See R1 Q; R3 W2; Sections 3.1–3.2, Eq. (7), Figure 2, Table 9, and Appendix A.3–A.5 / Algorithm 1.

  2. Attribution: how geometry is used vs. backbone, operator, or decoder (R2 W1/Q2–Q3, R3 W2/Q2).
    Shared requests asked for matched fusion alternatives and an explicit structural constraint distinguishing geometry-conditioned interaction from direct fusion. Under a frozen backbone / WDKE / GFDB / loss / schedule protocol that changes only the depth slot in HWPB, gated-drive C0 outperforms RGB self-drive (C1), depth-as-content (C2), and same-slot concat / cross-attention (C3/C4). The H0/H1H_0/H_1 split and equal-RMS high-band depth-noise results further align with the mechanism. See R2 Q2–Q3; R3 Q2; Section 4.4, Eq. (7), and Appendix Tables 28–29 / 33 (H.3, H.7).

  3. Robustness attribution and depth-path interpretation (R2 W2/Q1/Q3, R3 W2).
    Lower corruption sensitivity could have come from many factors. We now state that the original Table 4 compares complete training schemes; report a same-backbone PP-GG-AA factorial and within-method augmentation toggle; and separate useful-depth benefit UU from corruption sensitivity Dˉ\bar D with a frozen-weight probe. Table 6’s “w/o depth feature encoder” row is clarified as removing multi-scale depth feature learning rather than all geometric pathways, with full RGB-only GeoHaze reported separately. See R2 Q1/Q3–Q4; R3 W2; Sections 4.5–4.6 and Appendix Tables 27, 30–32, and 34 (H.2, H.4–H.6, H.8).

  4. RGB-only baselines, claim scoping, and evaluation protocol (R3 W1, R2 follow-up).
    Strong RGB-only methods belong in the comparison, and published PSNR numbers are not interchangeable with Table 1 under mismatched splits and inference settings. The revision modified the motivation to a conditional question about how imperfect geometry should enter the network; adds released-weight and matched-retraining comparisons with ConvIR, FocalNet, and MB-TaylorFormer V2; and corrects Appendix B.3 to state that I/O/NH-HAZE GeoHaze evaluations used 512×512512\times512 tiles, which is the same protocol used in the matched I/O-HAZE panel. See R3 W1; R2 protocol clarification; Sections 1 and 4.2, Figure 1, Appendix B.3, and Tables 23–27 (H.2).

  5. Efficiency and failure modes (R1 W1/W3).
    Beyond FLOPs, we report wall-clock latency and peak memory at 4K AMP against existing baselines, and expand the limitation discussion with qualitative failure cases for uniform dense haze, repetitive thin textures, and monocular-depth ambiguity. See R1 W1/W3; Table 8, Appendix Tables 20–22 (H.1), Section 4.6, and Appendix I / Fig. 14.

For the remaining reviewer-specific questions, the individual rebuttals add further comparisons, quantitative checks, ablations, and clarifications. In brief:

  1. Section 3 operator-domain wording and 4K efficiency / FFT cost (R1).
  2. Matched C0–C4 controls, H0/H1H_0/H_1 constraints, depth-reliance probes, and WDKE input roles (R2).
  3. Protocol-aligned RGB-only baselines, within-architecture depth-path gains, and WaveFormer novelty boundary (R3).
  4. Documentation correction for tiled inference on I/O/NH-HAZE (R2 follow-up; Appendix B.3).

We thank the reviewers for the careful feedback, and the Area Chairs for the time spent on this submission.

Authors of Submission 1433

I appreciate the authors' great devotion to this work, which undoubtedly involves a substantial amount of engineering effort. Nevertheless, I maintain a negative vote, as the innovation of this paper is limited

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.