{"id":"d4f32846-6764-4893-b338-d28dd65393f8","arxiv_id":"2506.15837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An adaptive, density-aware multi-stage dehazing framework routes foggy images to specialized restoration branches based on a learned fog-density score, improving dehazing quality and downstream object detection.","lead":"ADAM-Dehaze is a fog-removal pipeline that estimates fog density and then sends each image through one of three dedicated restoration networks. The authors report better image quality and better object detection in foggy scenes, but the paper contains several inconsistencies in the reported numbers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-world RTTS evaluation uses only the CORUN-Light branch, so the adaptive routing mechanism that is the paper's central contribution is never tested as a system.","rationale":"The paper has several useful building blocks: a synthetic dataset with three explicit β levels, a lightweight classifier with high synthetic accuracy, and an explicit design of three CORUN variants. These are real contributions even if the headline claims overstate them. However, the central proposition—that adaptivity based on a learned density score yields better restoration and detection than a uniform dehazer—is exactly the part of the paper that is never empirically closed. The reader focused on HDEN's transfer to real fog; I agree that is one necessary condition, but it is not sufficient. Even with perfect HDEN, the paper would still lack evidence because Table V evaluates a single branch, not the routed system, and no uniform or no-dehaze baseline is provided. The internal contradictions among the abstract (13-point mAP, 30% FADE, 20% faster), Table V (1-point mAP, 12% FADE), and Table VI (33% average latency) further weaken confidence in the reported numbers. Since these are addressable with additional experiments and code release, a conditional verdict remains appropriate; the concern does not prove the method fails, it shows the central claim is currently unverifiable.","tokens_in":9866,"tokens_out":4280,"duration_ms":47646,"concrete_test":"Run the released ADAM-Dehaze system (HDEN + thresholds + three CORUN branches) on the RTTS test set and report per-image routing decisions, FADE/BRISQUE/NIMA, YOLOv8 mAP, and total latency. Compare against (i) YOLOv8n on raw hazy images, (ii) a uniform 6-stage CORUN applied to all images, and (iii) each fixed branch applied to all images. If the full adaptive pipeline does not exceed the best single fixed branch on mAP and FADE by more than run-to-run noise, the adaptive-routing claim is unsupported. Additionally, inspect the routing distribution on RTTS: if nearly all real images are routed to one branch, the adaptive system reduces to a fixed model in practice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that density-based routing to specialized branches improves restoration and detection. The reported evidence never tests the adaptive system. Table V, the only real-world evaluation, explicitly says 'leveraging the CORUN-Light branch' and is labeled 'under light fog conditions'; no result is shown for the full HDEN+routing pipeline on RTTS. Consequently, even if HDEN is perfectly accurate, there is no demonstration that routing helps. The abstract's 'up to 13 points' mAP gain does not appear in Table V: the mean mAP is 0.56 vs. RIDCP's 0.55, a 1-point difference, and no no-dehazing YOLOv8n baseline is reported. On synthetic Cityscapes, Table IV lists PSNR/SSIM/LPIPS only for 'Ours', with no comparison method, so the claimed 'up to 2.1 dB' cannot be verified. The routing thresholds α and β in Eq. (4) and Algorithm 1 are never given, and HDEN's 99.80% accuracy is measured only on synthetic fog. The 20% inference-time saving in the abstract is inconsistent with Table VI's 33% average latency reduction, and no measured routing distribution on real fog is reported. Thus the load-bearing premise—that adaptive selection is accurate and beneficial in practice—has no direct evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADAM-Dehaze, an adaptive dehazing framework that first classifies a foggy image into light, medium, or heavy fog using a Haze Density Estimation Network (HDEN), then routes the image through one of three CORUN-based dehazing branches, and finally feeds the restored image into YOLOv8n for object detection. The authors introduce a synthetic FogIntensity-25K dataset and a density-modulated loss combining physical-model coherence, perceptual, and density terms. The claimed contributions are improved restoration (up to 2.1 dB PSNR, 30% FADE reduction) and detection (up to 13 mAP points) on Cityscapes and RTTS, with reduced inference time.","tokens_in":10155,"tokens_out":3101,"duration_ms":33362,"significance":"If the central claims were fully supported, the idea of routing images by estimated fog density to specialized restoration branches, while co-optimizing with a downstream detector, would be a useful engineering contribution to fog-robust perception. The paper has several strengths: the authors commit to releasing code, models, and the FogIntensity-25K dataset; they include ablations for the loss components and the cooperative proximal mapping module; and they evaluate on a real-world benchmark (RTTS) in addition to synthetic Cityscapes data. However, the evidence as presented does not yet support the load-bearing claims. The adaptive routing mechanism, which is the paper's main novelty, is never evaluated as a complete system on real foggy images, and the headline quantitative gains in the abstract are not consistently reflected in the tables. The central hypothesis remains plausible but unverified.","major_comments":[{"comment":"The only real-world detection evaluation is performed using the CORUN-Light branch only, as stated in the table caption ('Our method, leveraging the CORUN-Light branch') and in Section IV-E. This means the full adaptive system (HDEN routing plus all three branches) is never tested on RTTS. The abstract's claim of 'up to 13 points' mAP improvement is not visible in Table V: the mean mAP is 0.56 versus RIDCP's 0.55, a 1-point difference, and no no-dehazing YOLOv8n baseline is reported. The authors should report results for the full adaptive pipeline on RTTS, the routing distribution over the three branches, and a comparison against each fixed branch, in addition to the fixed-branch baseline.","section":"IV-E, Table V"},{"comment":"Table IV reports SSIM, PSNR, and LPIPS only for the proposed method, with no baseline rows, so the claimed 'up to 2.1 dB' PSNR improvement cannot be confirmed from the presented data. The text states that ADAM-Dehaze 'consistently achieves the lowest distortion scores and highest perceptual/structural metrics,' but no comparison values are shown. The authors should add the same metrics for DAD, MBDN, RIDCP, and PSD in Table IV, matching the comparison columns in Table III.","section":"IV-C-1, Table IV"},{"comment":"The routing thresholds α and β in Eq. (4) and Algorithm 1 are never specified, and the HDEN classification accuracy (99.80%, Table II) is measured only on 5,000 synthetic test images. Without explicit threshold values and without any real-world routing accuracy evaluation, the behavior of the adaptive system on RTTS is undefined. The authors should provide the chosen α and β values and report HDEN accuracy or routing statistics on real foggy images, including which branches RTTS images are routed to.","section":"III-B-2, Eq. (4), Algorithm 1"},{"comment":"The inference-time claims are inconsistent across the paper. The abstract states a 20% inference-time reduction; Section IV-F states a 33% average latency reduction and 23% FLOP savings; and the text below Table VI says '20–40%.' From Table VI, the adaptive average time (30 ms) versus the fixed 6-stage model (45 ms) gives a 33% reduction, not 20%. The authors should reconcile these numbers and report the actual measured routing distribution used to compute the average.","section":"IV-F, Table VI"},{"comment":"The density loss Ldens = ||D_theta_D(Jout)||_1 uses the same HDEN network that performs routing, but the paper does not state whether HDEN is frozen or fine-tuned during the joint training. If D_theta_D is optimized jointly with this loss, a trivial solution is for the density estimator to output zero on all dehazed images, making the loss vacuous without enforcing real haze removal. The training status of HDEN and any stop-gradient or fixed-weight treatment should be stated explicitly.","section":"III-B-3, Eq. (8)"}],"minor_comments":[{"comment":"The text attributes CORUN to reference [12], but [12] is the MSBDN paper by Dong et al. If CORUN is a separate method, it should have its own citation; if not, the description of 'cooperative unfolding modules' needs a correct reference.","section":"Section II, Related Work"},{"comment":"The coherence loss in Eq. (6) implicitly assumes global atmospheric light A = 1. This assumption is not stated in the text and is part of the physical-model coherence claim; it should be made explicit.","section":"Section IV-C-1, Eq. (6)"},{"comment":"The abstract claims a 30% FADE reduction, but Table V shows FADE of 0.828 for the proposed method versus 0.944 for RIDCP, which is a 12.3% reduction, and the text itself says '12.7% in FADE.' The 30% figure appears only in the abstract and in Section VI ('up to 30% lower FADE'). These numbers should be reconciled.","section":"Abstract and Section IV-C-2"},{"comment":"The ablation table reports mAP of 75.0% for the full model, while Table V reports mean mAP of 0.56 (56%) on RTTS. The differing scales and datasets are not explained; the ablation should state which dataset and protocol produce the 75.0% value.","section":"Table VII and Table V"},{"comment":"The claim in Section IV-E that 'these improvements are consistent across the different object categories' is only partially supported: the Bus category shows 0.31 for Ours versus 0.32 for RIDCP, so one category is actually lower. The wording should be softened or the per-category results discussed more carefully.","section":"IV-E, Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and timely problem, and the proposed framework has a plausible design. However, the current manuscript does not substantiate the central claim that adaptive routing improves real-world dehazing and detection: the only real-world evaluation uses a single branch, and the abstract's quantitative claims are not supported by the tables. These issues are fixable with additional experiments, but they require a substantial revision rather than minor edits. I would also encourage the editor to verify that the comparisons in Table IV are added and that the reported inference-time numbers are made internally consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read.\n\nThe core idea is sound: estimate fog density, route to a branch sized for that density, and co-tune with detection. That is a reasonable way to spend compute where it matters. The paper also ships a new synthetic dataset (FogIntensity-25K) and a clean ablation table. Credit is due for making the code link and for the three-branch CORUN assembly, which is genuinely a new arrangement.\n\nBut the evaluation does not back the headline claims. The synthetic result table (Table IV) has no baseline rows, so the \"up to 2.1 dB\" PSNR gain is unverifiable. On the real-world RTTS benchmark, Table V explicitly says the results come from the CORUN-Light branch only. That means the adaptive routing mechanism—the paper's central contribution—is never tested as a system on real fog. The abstract's 13-point mAP gain does not appear in the table: the mean mAP is 0.56 versus RIDCP's 0.55, about one point. There is no no-dehazing detector baseline, so we do not know what YOLOv8n alone scores on these images. The routing thresholds alpha and beta are never given. The 99.8% classification accuracy is only on synthetic fog, so HDEN's behavior on real RTTS images is unknown. And the abstract's 20% speed gain sits uneasily against Table VI's 33% average latency claim—not a big deal, but it signals loose reporting.\n\nThe density loss that uses HDEN to score the dehazed output is mildly self-referential, but it is not a circularity that sinks the method; it would need a check for trivial solutions.\n\nThese problems are addressable. The fix is a revised evaluation: add baselines to the full-reference table, run the full adaptive pipeline on RTTS and report which branches real images route to, add a no-dehazing detector baseline, and state the thresholds. If the authors can show that routing actually helps on real fog, this would be a useful engineering contribution for autonomous-driving perception.\n\nMy recommendation: send it to serious peer review, but with the expectation of major revision. The idea and dataset are worth refereeing; the claims as written are not.","headline":"Plausible adaptive-dehazing idea, but the paper never actually tests the adaptive routing on real fog, and the abstract's numbers don't match the tables.","tokens_in":10665,"tokens_out":2494,"would_cite":false,"duration_ms":28081,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that foggy images should be dehazed by a network selected according to the estimated fog density, and that doing so improves both restoration quality and downstream object detection compared with a single uniform…","keywords":["adaptive dehazing","haze density estimation","object detection in fog","atmospheric scattering model","deep unfolding network","multi-stage image restoration","foggy scenes","real-world dehazing"],"falsifier":"On real foggy images with visibly dense fog (for example, RTTS images of heavy haze), record which branch HDEN selects. If most are sent to the Light branch, or if using the predicted routing gives no better detection or restoration than forcing all images through one fixed branch, the adaptive advantage collapses. A cleaner control on synthetic data is to compare the full system against an oracle router that uses the ground-truth beta value; a small gap would indicate HDEN contributes little.","tokens_in":9648,"feed_emoji":"🌫️","tokens_out":7072,"duration_ms":66628,"temperature":0.7,"pith_summary":"The paper tries to show that fog removal is best done adaptively, with the network capacity and loss weighting matched to the estimated fog density. It proposes a framework that first computes a haze density score for each image, then routes the image to one of three dehazing networks designed for light, medium, or heavy fog. On synthetic Cityscapes and real-world RTTS images, the authors report PSNR gains of up to 2.1 dB, a 30% reduction in FADE, object detection mAP gains of up to 13 points, and a 20% reduction in inference time. If correct, this would mean that density-aware routing is a practical way to make dehazing genuinely useful for downstream tasks like autonomous driving.","feed_headline":"Fog density-aware dehazing lifts detection mAP by 13 points","feed_subtitle":"Three branches tuned to light, medium, and heavy fog restore images and cut inference time by 20 percent.","key_machinery":"The load-bearing mechanism is the coupling between the Haze Density Estimation Network (HDEN) and the three CORUN branches. CORUN is a cooperative unfolding network that alternates steps estimating scene radiance and transmission; the paper instantiates it with 2, 4, and 6 stages, the heaviest branch adding transformer-based attention. HDEN's predicted score $d$ determines not only which branch runs, but also the mix of loss terms via $\\gamma(d) \\in \\{0.3, 0.6, 0.9\\}$. This coupling makes the network capacity and the optimization objective both vary with fog density, which is the backbone of the claimed gains.","core_discovery":"The paper claims to show that a density-aware, multi-branch system beats a monolithic dehazer on both image quality and detection. A lightweight Haze Density Estimation Network (HDEN) outputs a continuous score $d \\in [0,1]$; thresholds $\\alpha$ and $\\beta$ route the input to a 2-stage CORUN-Light, 4-stage CORUN-Medium, or 6-stage CORUN-Complex branch. The same score sets the loss weight $\\gamma$ between physical-coherence, perceptual, and density terms. On synthetic Cityscapes the authors report PSNR up to 23.95 dB, SSIM up to 0.9188, and LPIPS down to 0.0585, and on real-world RTTS they report FADE 0.828, BRISQUE 11.961, NIMA 5.346, and mean detection mAP 0.56, with detection improvements up to 13 mAP points over hazy inputs.","pith_inferences":["A direct test: compare predicted routing against oracle routing on real fog images. If the oracle does not beat the learned router, HDEN's classification accuracy is not what drives the reported gains.","The routing in Equation (4) is a hard switch; a soft mixture of branch outputs near the thresholds might behave better on continuous haze densities and on images that straddle two regimes.","Because HDEN is trained on three discrete synthetic beta values, real fog with spatially varying density and arbitrary atmospheric light creates a distribution gap; measuring the router's confidence on out-of-distribution fog would quantify this.","If the density-aware principle holds, the same router-plus-branch design could be transferred to other degradations such as rain, snow, or low light, each with its own cheap and expensive restoration paths."],"forward_implications":["In light fog the 2-stage branch runs at 18 ms per image, so the adaptive scheme is fast enough for real-time use in vehicles while reserving the 50 ms 6-stage branch for dense fog.","Because the dehazing branches are fine-tuned together with a detector, the reported mAP gains come from features aligned to detection, not only from cleaner pixels.","The density-modulated loss assigns higher weight to physical-model coherence in heavy fog, implying that physics-based constraints matter most precisely where pure learning struggles.","The synthetic FogIntensity-25K dataset with $\\beta = 0.03, 0.06, 0.09$ gives the community paired hazy/clear images with depth-aware fog, enabling full-reference evaluation across intensity levels."],"supporting_citations":[{"why":"Supplies the Cityscapes clear images and depth maps used to synthesize the FogIntensity-25K training and evaluation pairs.","marker":"[23]"},{"why":"Supplies the Synscapes photorealistic street scenes used to build the foggy synthetic data.","marker":"[37]"},{"why":"Provides the RTTS real-world foggy benchmark and the FADE evaluation protocol used for non-reference dehazing assessment.","marker":"[3]"},{"why":"Defines the dark channel prior and the atmospheric-scattering-model baseline that the physical coherence loss builds on.","marker":"[2]"},{"why":"Cited as the CORUN cooperative unfolding base whose stage count the three branches scale from 2 to 6 stages.","marker":"[12]"},{"why":"The PSD baseline, a density-guided physical-prior dehazing method that ADAM-Dehaze compares against and must beat on restoration and detection.","marker":"[19]"},{"why":"The RIDCP baseline using high-quality codebook priors, the strongest image-quality competitor on RTTS in the paper's tables.","marker":"[18]"},{"why":"Prior work on unified density-aware dehazing and detection that this paper extends by adding explicit routing and adaptive loss weighting.","marker":"[16]"},{"why":"Defines the FADE fog-density metric used to measure residual haze on real images without ground truth.","marker":"[34]"}],"fun_headline_variants":["Fog density-aware dehazing adds 13 mAP, cuts runtime 20%","Adaptive multi-branch dehazing lifts foggy mAP by 13","Density-routed dehazing: 13 point mAP boost for fog","HDEN-guided dehazing improves detection mAP by 13","Three-branch dehazing tunes fog density, gains 13 mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that fog synthesized with the atmospheric scattering model at three fixed density values ($\\beta = 0.03, 0.06, 0.09$) teaches the density estimator to route real-world foggy images correctly.","fun_headline_variants_meta":{"raw":{"variants":["Fog density-aware dehazing adds 13 mAP, cuts runtime 20%","Adaptive multi-branch dehazing lifts foggy mAP by 13","Density-routed dehazing: 13 point mAP boost for fog","HDEN-guided dehazing improves detection mAP by 13","Three-branch dehazing tunes fog density, gains 13 mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2950,"prompt_tokens":966,"completion_tokens":1984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":1885}},"tokens_in":582,"tokens_out":1984,"duration_ms":16305,"temperature":1.0,"reasoning_tokens":1885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:37:00.951071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On real foggy images with visibly dense fog (for example, RTTS images of heavy haze), record which branch HDEN selects. If most are sent to the Light branch, or if using the predicted routing gives no better detection or restoration than forcing all images through one fixed branch, the adaptive advantage collapses. A cleaner control on synthetic data is to compare the full system against an oracle router that uses the ground-truth beta value; a small gap would indicate HDEN contributes little.","supporting_citations":[{"cited_title":"Benchmarking single-image dehazing and beyond,","cited_arxiv_id":null,"evidence_quote":"Provides the RTTS real-world foggy benchmark and the FADE evaluation protocol used for non-reference dehazing assessment."},{"cited_title":"Single image haze removal using dark channel prior,","cited_arxiv_id":null,"evidence_quote":"Defines the dark channel prior and the atmospheric-scattering-model baseline that the physical coherence loss builds on."},{"cited_title":"Multi-scale boosted dehazing network with dense feature fusion,","cited_arxiv_id":null,"evidence_quote":"Cited as the CORUN cooperative unfolding base whose stage count the three branches scale from 2 to 6 stages."},{"cited_title":"Psd: Principled synthetic-to-real dehazing guided by physical priors,","cited_arxiv_id":null,"evidence_quote":"The PSD baseline, a density-guided physical-prior dehazing method that ADAM-Dehaze compares against and must beat on restoration and detection."},{"cited_title":"Ridcp: Revitalizing real image dehazing via high-quality codebook priors,","cited_arxiv_id":null,"evidence_quote":"The RIDCP baseline using high-quality codebook priors, the strongest image-quality competitor on RTTS in the paper's tables."},{"cited_title":"Unified density-aware image dehazing and object detection in real- world hazy scenes,","cited_arxiv_id":null,"evidence_quote":"Prior work on unified density-aware dehazing and detection that this paper extends by adding explicit routing and adaptive loss weighting."},{"cited_title":"Referenceless prediction of perceptual fog density and perceptual image defogging,","cited_arxiv_id":null,"evidence_quote":"Defines the FADE fog-density metric used to measure residual haze on real images without ground truth."}],"review_version":1}