{"id":"e0a671e3-2568-4aa1-ad56-851329622982","arxiv_id":"2412.11395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A joint neural network trained with atmospheric-scattering and brightness-consistency losses simultaneously dehazes real driving video and estimates depth, reporting state-of-the-art results on four hazy benchmarks.","lead":"Researchers built a system that removes haze from driving videos while also estimating scene depth in one pass, training it on real foggy footage rather than synthetic pairs. The paper reports better results than current methods on several real-world hazy benchmarks, which matters for autonomous driving and video safety systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BCC rigidity assumption in Eq. (3) is untested on dynamic driving scenes; the auto-mask does not remove motion-corrupted gradients, so the joint-training claim rests on an assumption that is false in the target domain.","rationale":"The paper's central claim is that the BCC-ASM sharing loop is what makes joint self-supervised training on real hazy driving video work. Ablations (Table 3: removing BCC raises DENSE-Fog light Abs Rel from 0.311 to 0.636) provide real evidence that BCC contributes, and the DMFIR/DMDR/NRFM components add plausible regularization. But the BCC gradient is only as trustworthy as the rigidity assumption in Eq. (3). The auto-mask (Eq. 7) is inherited from Monodepth2 and is designed to handle occlusion and stationary-object appearance change, not to segment moving objects; in homogeneous regions of moving vehicles it can fail open and pass wrong gradients. The paper's own limitation note says dynamic scenes with people and vehicles are hard to obtain, but it never tests the model under those conditions. This is exactly the gap between works on the submitted datasets and works for real-world hazy driving video. A separate red flag in Table 2 is the dense-fog Abs Rel of 1.182 versus 0.605 for RobustDepth, which already weakens the absolute SOTA claim; it does not, however, identify the mechanism, whereas the rigidity violation does. The reader's weakest assumption was the same static-scene concern, so agreement is agree; since the appropriate response is additional evidence rather than outright rejection, the verdict stays CONDITIONAL/UNCHANGED.","tokens_in":16673,"tokens_out":8707,"duration_ms":84916,"concrete_test":"Retrain DCL with Eq. (6) multiplied by a static-only mask from Mask2Former segmentation of vehicles, pedestrians, and cyclists on GoProHazy and DrivingHazy, then compare against the published DCL on DENSE-Fog metrics and on a dynamic-object holdout with sparse LiDAR. If dynamic-region depth error drops while static-region metrics are unchanged, the unfiltered BCC is injecting motion-corrupted gradients; if metrics are unchanged, the rigidity concern is not load-bearing. Also report auto-mask activation rates on dynamic versus static pixels from the trained checkpoint.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing pillar of the method is Eq. (3): the BCC branch assumes that every pixel in frame t is related to frame s by a single rigid pose, y ~ K P_{x->y} d(x) K^{-1} x, with d predicted by the shared depth network. In the target application (real hazy driving video), this rigidity is violated by moving vehicles and pedestrians. The auto-mask in Eq. (7) only suppresses pixels where the photometric error is not reduced by the warp; it does not detect or exclude dynamic objects. For low-texture moving surfaces, the wrong warp can still achieve a low Lpe, so erroneous gradients propagate into both the depth network and the dehazing network through the shared ASM term. The paper reports no analysis, ablation, or metric that isolates dynamic regions; the supplement's limitation note mentions difficulty collecting data in dynamic scenes but does not examine the model's behavior when such objects are present. Because BCC is the mechanism claimed to make joint training beneficial, the central claim that self-supervised real hazy driving video is a viable route rests on an assumption that is known false in the domain and untested here. This does not invalidate the empirical results on the evaluated mostly static sequences, but it makes the real-world driving generalization claim unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Depth-Centric Learning (DCL), a self-supervised framework for joint video dehazing and monocular depth estimation from real hazy driving videos. The central idea is a unified ASM-BCC model in which a shared depth network feeds both the atmospheric scattering model (ASM), used to reconstruct the hazy frame from predicted depth, scattering coefficient, airlight, and a dehazed frame, and a brightness consistency constraint (BCC), which reprojects dehazed neighboring frames via predicted depth and pose. Because aligned ground truth for real hazy driving video is unavailable, the method uses misaligned clear reference videos with a non-aligned reference frame matching loss and two discriminators: DMFIR for high-frequency detail preservation and DMDR for reducing black-hole artifacts in weak-texture depth. Experiments are reported on GoProHazy, DrivingHazy, InternetHazy, and DENSE-Fog, with a supplementary REVIDE experiment providing an aligned-ground-truth sanity check. The paper claims state-of-the-art performance in both dehazing and depth estimation, with particularly large margins on no-reference dehazing metrics.","tokens_in":16927,"tokens_out":8863,"duration_ms":79019,"significance":"If the claimed results hold, the paper makes a useful contribution: it demonstrates that a physical-model-based self-supervised loop can jointly dehaze and estimate depth from real hazy video, and the two regularizers (DMFIR and DMDR) address real artifacts without paired supervision. The method is efficient (low inference time and parameter count), and the publicly available project page and the REVIDE aligned-GT experiment are concrete strengths that go beyond no-reference evaluation. However, two concerns limit full confidence: the BCC branch in Eq. (3) assumes a static rigid scene, which is violated in the target driving domain and is not analyzed; and the dense-fog depth results in Table 2 are inconsistent with the paper's overall claim of superiority. The ablations in Table 3 and Table 4 do support the importance of BCC, DMFIR, DMDR, and the losses, which is reassuring, but the dense-fog anomaly and the lack of motion handling prevent an unconditional recommendation.","major_comments":[{"comment":"The BCC branch assumes that every pixel in frame t is related to frame s by a single rigid pose, i.e., y ~ K P_{x->y} d(x) K^{-1} x. In real driving videos, vehicles and pedestrians violate this assumption, and the auto-mask in Eq. (7), taken from Monodepth2, only suppresses pixels where warping does not reduce the photometric error; it is not a motion-segmentation or scene-flow mechanism. The manuscript provides no experiment or analysis isolating dynamic regions, and the supplement's Limitations section explicitly states that collecting high-quality misaligned data in dynamic scenes is challenging. Because the BCC is the mechanism claimed to make joint training beneficial, this untested static-scene assumption is load-bearing for the paper's real-world driving generalization claim. Please either incorporate a motion mask or instance segmentation to exclude dynamic objects, or provide an evaluation (e.g., using DENSE-Fog's radar ground truth in regions flagged as dynamic, or an object-detector-based split) showing that moving objects do not corrupt the depth and dehazing gradients.","section":"A Unified ASM-BCC Model, Eqs. (3), (6)-(7)"},{"comment":"On the DENSE-Fog dense split, DCL reports the worst Abs Rel (1.182) among all compared methods (e.g., RobustDepth 0.605, Mono-ViFI 0.609, Lite-Mono 0.954) and its RMSE log (0.596) is also worse than RobustDepth (0.515) and Mono-ViFI (0.528). The sentence in Section 'Monocular depth estimation' that DCL 'outperforms the others across nearly all five evaluation metrics' is only true for the light-fog split; the dense-fog results directly contradict the broader claim that DCL excels in real hazy scenes. The parenthetical explanation that 'blurred depth estimates tend to be closer to the mean of the ground truth' is not quantified and does not justify the Abs Rel regression. Please either restrict the claim to light-fog conditions, provide an analysis of why dense-fog depth fails (e.g., predicted depth histograms vs. ground truth), or report per-sequence results to identify whether the failure is systematic or due to a few outliers.","section":"Experiment Results, Table 2"},{"comment":"The main dehazing comparisons on GoProHazy, DrivingHazy, and InternetHazy rely exclusively on no-reference metrics (FADE and NIQE) without error bars, repeated runs, or statistical significance tests. These metrics are known to be sensitive to contrast and color statistics, so the reported improvements (e.g., NIQE 3.4412 vs. DVD 4.4473) might not reflect perceptually meaningful gains. The supplementary REVIDE experiment provides aligned ground truth, but on indoor smoke, not driving haze. If aligned driving data is genuinely unavailable, the authors should report standard deviations over multiple training runs, or add a qualitative user study, or evaluate on an additional real hazy driving dataset with post-hoc alignment, to substantiate the claim of state-of-the-art real-world dehazing.","section":"Datasets and Evaluation Metrics, Table 1"}],"minor_comments":[{"comment":"The symbol '~' is used to denote pixel correspondence, but '~' conventionally means 'distributed as' in statistics; please replace it with a clear arrow or the phrase 'is mapped to' to avoid ambiguity.","section":"Eqs. (2) and (3)"},{"comment":"The symbol 'S' is used both for the differentiable bilinear sampling operation and for the SSIM loss; rename one of them (e.g., use 'B' for bilinear sampling) to prevent confusion.","section":"Eqs. (3) and (6)"},{"comment":"The bracket notation [Lpe(Jt, bJt) < Lpe(Jt, Js)] should be explicitly defined as an indicator function, and the text should clarify that ma is applied element-wise to the photometric loss in Eq. (11).","section":"Eq. (7)"},{"comment":"The claim 'To the best of our knowledge, we are the first' is strong; consider softening to 'To our knowledge' and citing any recent concurrent works on joint depth estimation and dehazing from real video.","section":"Introduction, Contributions"},{"comment":"The table repeats the five metric headers for light-fog and dense-fog without a clear visual separation; please add subheadings or a blank column so the two splits are immediately distinguishable.","section":"Table 2"},{"comment":"The limitations paragraph (dynamic scenes, small object detail) is placed only in the supplement; please refer to it explicitly in the main text, since it directly qualifies the real-world generalization claim.","section":"Supplemental Material, Limitations"},{"comment":"The 'Discussion on predicted depth surpassing reference depth' (Fig. S4) is based on a single visual example; if this claim is important, provide a quantitative comparison (e.g., mean Abs Rel on a subset) between the predicted depth and MonoDepth2 reference depth.","section":"Supplemental Material, Sec. D"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its core self-supervised formulation, and the REVIDE experiment is a commendable addition. However, the dense-fog depth result in Table 2 is a clear inconsistency that the authors must address or explicitly caveat. The static-scene assumption in the BCC branch is a known gap in self-supervised depth estimation, and the authors' own limitation note admits it; a revision that adds a motion-handling mechanism or a careful analysis of dynamic regions would substantially strengthen the paper. The novelty over DVD is mostly in the integration of ASM and BCC with discriminators, which is reasonable but should be positioned precisely."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the DCL paper (arXiv:2412.11395). Short take: it's a plausible new combination of video dehazing and self-supervised depth, built from the authors' own DVD work, and the ablations are clean, but the central joint-training loop leans on a static-scene assumption that real driving video violates, and the dense-fog depth numbers are weird enough that I wouldn't take the headline claims at face value.\n\nWhat's new: the unified ASM-BCC model, which uses a shared depth network to tie atmospheric-scattering reconstruction to brightness-consistency reprojection between adjacent dehazed frames, is a sensible integration I haven't seen in that exact form. The two discriminator regularizers (DMFIR for high-frequency detail, DMDR for depth black holes) are also new, and the idea of pulling reference depth from a non-aligned clear video to regularize hazy-depth estimation is neat. The writing is clear, and the ablation study actually supports the design: removing BCC, DMFIR, or DMDR each hurts, and the REVIDE experiment provides a good aligned-ground-truth check. Efficiency is fine.\n\nSoft spots, in order of concern. First, the BCC branch assumes rigid scene geometry (Eq. 3) with a single pose per frame; the auto-mask (Eq. 7) suppresses pixels where the photometric error doesn't improve, but that does not reliably exclude moving vehicles or pedestrians, especially low-texture ones. The supplement's limitation note admits dynamic data is hard to collect but never examines what the model does when dynamic objects are present. That's a real gap for a driving application. Second, Table 2 shows the dense-fog Abs Rel for DCL is 1.182, the worst of all compared methods, while other metrics look fine; the text's explanation about blurred depths being closer to the mean is hand-wavy and doesn't inspire confidence. Third, the main dehazing comparisons use only no-reference metrics FADE and NIQE, with no variance or statistical testing. Fourth, the paper cites joint dehazing+depth works (Li et al. 2015, DEHRformer) but doesn't compare to them. No code is released, so reproducibility rests on the project page.\n\nNone of this is fatal: the core idea is reasonable and the evidence on mostly static sequences is consistent. But the generalization claim to real driving scenes is not yet supported. I'd send it to serious review, with a request for analysis of dynamic regions, a clear explanation or fix for the dense-fog anomaly, and code release.","headline":"Plausible new joint dehazing/depth framework, but the static-scene assumption in the BCC loop and the dense-fog depth anomaly keep it from being fully convincing.","tokens_in":17484,"tokens_out":2370,"would_cite":false,"duration_ms":22101,"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 claims that one shared per-pixel depth map can drive both haze removal and monocular depth estimation in real hazy driving videos, with the two tasks reinforcing each other and beating separate state-of-the-art methods on both.","keywords":["video dehazing","depth estimation","atmospheric scattering model","brightness consistency constraint","self-supervised learning","real hazy driving videos","adversarial regularization","monocular depth"],"falsifier":"Render or record a hazy driving sequence with known per-pixel depth and a clearly moving foreground object, then run the model with and without the brightness-consistency branch while measuring depth error only inside the moving object's mask. If the moving region's depth error is systematically worse with the branch enabled, the static-scene assumption is the failure mode and the photometric gradient is corrupting the shared depth estimate.","tokens_in":1834,"feed_emoji":"🌫️","tokens_out":1779,"duration_ms":91381,"temperature":0.7,"pith_summary":"This paper is trying to establish that haze removal and monocular depth estimation from a moving camera can be solved as one problem rather than two. The proposal is a Depth-Centric Learning framework in which a single depth network feeds both the atmospheric scattering model, which reconstructs the hazy frame from a dehazed estimate, and the brightness consistency constraint, which matches pixels between adjacent dehazed frames. Because the same per-pixel depth appears in both models, cleaner dehazed frames sharpen the depth estimate and sharper depth improves the haze removal, creating a mutually reinforcing loop that can be trained on real hazy driving video without aligned ground truth. The authors report that this joint model beats the previous best methods on real benchmarks, for example lowering the no-reference haze quality score NIQE from 4.45 to 3.44 on GoProHazy and the relative depth error Abs Rel from 0.369 to 0.311 on DENSE-Fog light haze.","feed_headline":"Dehazing and depth in foggy driving video, one shared depth map","feed_subtitle":"One depth map links haze physics to frame-to-frame matching, beating separate methods on real foggy driving video.","key_machinery":"The engine of the paper is the unified ASM-BCC model of Eq. (3), which writes the hazy frame as $I_t(x)=J_t(x)e^{-\\beta d(x)}+A_\\infty(1-e^{-\\beta d(x)})$ and simultaneously writes the dehazed target frame as a reprojection $J_t(x)=S(J_s,y)$ with $y\\sim K P_{x\\to y}d(x)K^{-1}x$. Here $d$ is the single per-pixel depth map from the shared depth network, $J_t$ and $J_s$ are dehazed frames, $\\beta$ is a learned non-uniform scattering coefficient, $P_{x\\to y}$ is the relative pose, and $S$ is differentiable bilinear sampling. The depth map is the pivot: it is the variable through which haze physics and temporal geometry exchange information. Two regularizers around this core, a wavelet-frequency discriminator that restores high-frequency detail in dehazed frames and a mean-normalized depth discriminator that suppresses black holes in low-texture regions, use a non-aligned clear video and its estimated depth as supervision.","core_discovery":"The central discovery is that the atmospheric scattering model and the brightness consistency constraint can be unified by making them both depend on the same depth map, and that this shared dependency is what lets the two tasks improve each other. Given a current hazy frame and its neighbors, the model predicts a dehazed pair, a per-pixel depth $d$, a spatially varying scattering coefficient $\\beta$, and a relative pose; the scattering model then reconstructs the hazy frame from the dehazed estimate and depth, while the photometric loss reprojects pixels between the dehazed frames using that same $d$. The paper argues that this coupling, regularized by two discriminators that compare the dehazed video and the estimated depth against a non-aligned clear video, is enough to train directly on real hazy driving footage and that it produces both cleaner frames and more reliable depth than running either task separately.","pith_inferences":["Extrapolating the paper's architecture, the same shared-depth loop could be applied to other weather degradations such as rain, snow, or low light by swapping the ASM branch for the corresponding forward degradation model; the paper does not test this.","The static-scene assumption is the most plausible failure point: extending the auto-mask with an explicit motion-segmentation head, and evaluating depth error separately on dynamic and static regions of a hazy benchmark, would test whether moving objects are the main source of residual error.","Because the non-uniform $\\beta$ network is free to vary per pixel, it could absorb misestimation of depth or airlight; constraining $\\beta$ with a haze-density or polarization prior would reveal whether the reported depth gains come from true scattering estimation or from an overly flexible parameter.","The paper's observation that predicted depth visually surpasses the reference depth obtained from a clear-video depth network suggests the ASM reconstruction loss acts as an extra depth cue beyond photometric reprojection; this could be verified by ablating the reconstruction loss and measuring depth error alone."],"forward_implications":["Dehazing and depth estimation can be trained together on real, unaligned hazy driving video, removing the need for synthetic hazy/clear pairs or aligned depth ground truth.","A single per-pixel depth map is the coupling variable: cleaner depth improves haze removal through the scattering model, and cleaner dehazed frames improve depth through reprojection, so gains in either task reinforce the other.","A spatially varying scattering coefficient $\\beta$, rather than a constant one, is needed for real patchy haze and measurably improves depth estimation accuracy.","The coupled model also works on real indoor smoke data, where it reaches the highest PSNR and SSIM among the compared methods.","At test time the dehazing and depth branches can be detached and run separately, with inference times of 0.075 s and 0.009 s at $640\\times192$."],"supporting_citations":[{"why":"Supplies the atmospheric scattering model that the ASM branch uses to reconstruct hazy frames from depth and dehazed estimates.","marker":"(McCartney 1976)"},{"why":"Formalizes vision through participating media and the depth-dependent transmission $t=e^{-\\beta d}$ that connects depth to haze.","marker":"(Narasimhan and Nayar 2002)"},{"why":"Provides the dark channel prior used to estimate the airlight $A_\\infty$.","marker":"(He, Sun, and Tang 2010)"},{"why":"Provides the photometric error formulation and differentiable bilinear sampling used in the brightness consistency loss.","marker":"(Godard, Mac Aodha, and Brostow 2017)"},{"why":"Supplies the self-supervised depth pretraining and the auto-mask strategy that suppresses invalid photometric pixels in the BCC loss.","marker":"(Godard et al. 2019)"},{"why":"Introduces the non-aligned reference frames matching strategy and the GoProHazy, DrivingHazy, and InternetHazy datasets used for dehazing evaluation.","marker":"(Fan et al. 2024)"},{"why":"Provides the DENSE-Fog dataset with sparse radar depth ground truth used for depth evaluation in real hazy scenes.","marker":"(Bijelic et al. 2020)"},{"why":"Supplies the wavelet transform technique used in the DMFIR frequency-domain discriminator.","marker":"(Gao et al. 2021)"}],"fun_headline_variants":["One shared depth map removes haze and estimates distance in driving videos","Depth-centric coupling clears fog and yields depth from monocular video","Foggy driving video: one depth map to dehaze and measure distance","Shared depth map links haze removal and distance estimation in real fog","Dehaze and depth from one shared depth map in hazy driving videos"],"cache_read_input_tokens":19584,"weakest_assumption_plain":"The load-bearing premise is that the scene between adjacent frames is static and rigid, so every pixel's motion is explained by one camera pose and its own depth; when vehicles or pedestrians move, that link breaks and the brightness-consistency loss sends wrong gradients into the shared depth and dehazing networks.","fun_headline_variants_meta":{"raw":{"variants":["One shared depth map removes haze and estimates distance in driving videos","Depth-centric coupling clears fog and yields depth from monocular video","Foggy driving video: one depth map to dehaze and measure distance","Shared depth map links haze removal and distance estimation in real fog","Dehaze and depth from one shared depth map in hazy driving videos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2333,"prompt_tokens":959,"completion_tokens":1374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1282}},"tokens_in":575,"tokens_out":1374,"duration_ms":9712,"temperature":1.0,"reasoning_tokens":1282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:58:44.670967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render or record a hazy driving sequence with known per-pixel depth and a clearly moving foreground object, then run the model with and without the brightness-consistency branch while measuring depth error only inside the moving object's mask. If the moving region's depth error is systematically worse with the branch enabled, the static-scene assumption is the failure mode and the photometric gradient is corrupting the shared depth estimate.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the atmospheric scattering model that the ASM branch uses to reconstruct hazy frames from depth and dehazed estimates."},{"cited_title":"G.; and Nayar, S","cited_arxiv_id":null,"evidence_quote":"Formalizes vision through participating media and the depth-dependent transmission $t=e^{-\\beta d}$ that connects depth to haze."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the photometric error formulation and differentiable bilinear sampling used in the brightness consistency loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised depth pretraining and the auto-mask strategy that suppresses invalid photometric pixels in the BCC loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the non-aligned reference frames matching strategy and the GoProHazy, DrivingHazy, and InternetHazy datasets used for dehazing evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DENSE-Fog dataset with sparse radar depth ground truth used for depth evaluation in real hazy scenes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet transform technique used in the DMFIR frequency-domain discriminator."}],"review_version":1}