{"id":"9797a73e-3fee-4956-8d74-f5706e2f0123","arxiv_id":"2502.02027","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Applying a fog-trained, attention-guided dehazing step inside a YOLO detection cascade improves foggy mAP but degrades clear-image mAP, reversing the usual clear-to-foggy performance drop.","lead":"A cascaded pipeline that detects regions, dehazes them with a spatial-attention-enhanced AOD-Net, then re-detects, improves foggy-image detection but lowers accuracy on clear images. The paper reports this inverse trend and argues that selective preprocessing is needed instead of universal enhancement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The clear-image penalty in Table 1 is not causally tied to selective attention: no ablation compares AOD-NetX with and without the RoI mask, and Appendix F.1 contradicts Table 1 by saying dehazing consistently enhances clear detection.","rationale":"The reader's weakest assumption identifies exactly the confound I consider load-bearing: the clear-image degradation is attributed to selective cascaded dehazing without isolating the RoI attention module from the fact that AOD-NetX was trained only on fog. My reading adds two concrete supports: the missing all-ones-mask ablation is the minimal control that would separate these factors, and the manuscript contains an internal inconsistency—Appendix F.1 claims dehazing consistently improves clear images while Table 1 shows AOD-NetX reducing clear mAP below the no-dehazing baseline. That inconsistency raises the stakes for the proposed control: if the all-ones AOD-NetX also degrades clear mAP, the headline 'from fog to failure' would be an artifact of domain mismatch rather than a property of selective preprocessing; if it does not, the paper would have a solid negative-result contribution. The reader's conditional verdict is appropriate because the anomaly is plausible and consistently signed across YOLOv5 and YOLOv8, but the causal claim is not yet established. I would not change the verdict without seeing the ablation, so UNCHANGED is the right recommendation.","tokens_in":11539,"tokens_out":4036,"duration_ms":42636,"concrete_test":"Replace the RoI-derived spatial attention mask in AOD-NetX with an all-ones mask, keeping the dehazing network, training, and YOLOv5x detector fixed, and rerun the clear Cityscapes evaluation reported in Table 1. If all-ones AOD-NetX also lowers clear mAP to about 0.49, the penalty is caused by applying a fog-trained dehazer to clear inputs, not by selective masking; if all-ones AOD-NetX keeps clear mAP near 0.5644, the RoI mask is the cause. Report per-image paired mAP and multiple seeds so the 0.4896-vs-0.5644 gap is shown to be outside noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that selective cascaded dehazing degrades clear-image detection—rests on an untested comparison. Table 1 contrasts global AOD-Net + YOLOv5x (clear mAP 0.6813) with YOLOv5s + AOD-NetX + YOLOv5x (clear mAP 0.4896), but these pipelines differ in two confounded ways: the dehazing architecture (AOD-NetX adds a spatial attention module) and the use of RoI masking. AOD-NetX is trained only on Foggy Cityscapes, and Table 3 shows it behaves anomalously on out-of-distribution inputs (SSIM 0.656 on RTTS vs 0.932 for AOD-Net), so applying it to clear Cityscapes could itself introduce distortions. The paper's own Appendix F.1 states that 'the integration of dehazing modules, such as AOD-Net and AOD-NetX, consistently enhances object detection in both clear and foggy conditions,' directly contradicting Table 1's clear-image drop to 0.4896. Without a control that keeps AOD-NetX fixed and removes only the RoI-derived spatial attention mask, the observed inverse trend cannot be attributed to selective region enhancement rather than to feeding clear images through a fog-trained dehazer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-stage detection pipeline (\"Perceptual Piercing\") in which a lightweight detector identifies RoIs, AOD-NetX (a spatial-attention-enhanced dehazer) processes only those regions, and a heavy detector performs final recognition. The headline empirical claim, in Section 4 and Table 1, is that this selective pipeline exhibits an inverse trend: it improves mAP on foggy images relative to clear images (e.g., +25.68% for YOLOv5s+AOD-NetX+YOLOv5x), whereas conventional detectors and global AOD-Net integration show the expected clear-to-foggy drop. The paper interprets this as an implicit domain adaptation effect and discusses trade-offs of cascaded enhancement. However, the evidence for the causal role of selective region enhancement is incomplete, and the manuscript contains a direct internal contradiction between Table 1/Table 4 and the textual claim in Appendix F.1. The out-of-distribution results in Table 5 also weaken the \"effective in foggy conditions\" framing.","tokens_in":11769,"tokens_out":2778,"duration_ms":28528,"significance":"If the inverse trend is robust, the paper points to a genuinely surprising interaction between selective dehazing and object detection, with implications for autonomous driving and other vision pipelines that chain image restoration with high-level tasks. The paper ships an open implementation and evaluates across two YOLO families, three dehazing backbones, and multiple datasets, which is a useful empirical contribution. Yet the central causal interpretation is currently underdetermined: the AOD-NetX pipeline differs from the AOD-Net baselines in both architecture and RoI masking, and no ablation isolates the selective attention mechanism. The internal contradiction in Appendix F.1 and the OOD results in Table 5 are serious enough that the paper's main conclusion cannot be accepted as stated.","major_comments":[{"comment":"The central claim that selective region enhancement degrades clear-image detection is not supported by the presented comparisons. The AOD-Net+YOLOv5x row (clear mAP 0.6813) and the YOLOv5s+AOD-NetX+YOLOv5x row (clear mAP 0.4896) differ in two confounded ways: the dehazing architecture (AOD-NetX adds a spatial attention module) and the RoI-masked application of dehazing, in addition to the lightweight detector stage. No ablation keeps AOD-NetX fixed and removes only the RoI-derived attention mask. Without such a control, the clear-image penalty cannot be attributed to selective region enhancement rather than to applying a fog-trained dehazer (with its modified architecture) to clear inputs. Please add this ablation or explicitly reframe the conclusion as a property of the full pipeline, not of selectivity.","section":"Section 4, Table 1; Section 5; Appendix D.4"},{"comment":"There is a direct internal contradiction: Appendix F.1 states that \"the integration of dehazing modules, such as AOD-Net and AOD-NetX, consistently enhances object detection in both clear and foggy conditions,\" but Table 4 shows that AOD-NetX integration reduces clear mAP from 0.5644 to 0.4896 for YOLOv5x and from 0.5243 to 0.5150 for YOLOv8x. This is not a minor wording issue; it undermines the paper's explanation of the inverse trend. Please correct the text and clarify whether AOD-NetX has a net positive or negative effect on clear images.","section":"Appendix F.1, Table 4"},{"comment":"The out-of-distribution evaluation contradicts the unqualified claim that the proposed pipeline is \"effective in foggy conditions.\" On RESIDE-β RTTS, YOLOv5s+AOD-NetX+YOLOv5x achieves mAP 0.5297 versus 0.6655 for the baseline YOLOv5x, and the AOD-NetX variants are consistently below their baselines on both OTS and RTTS. If the claim of foggy-domain effectiveness is meant only for the in-distribution Foggy Cityscapes benchmark, that scope limitation must be stated prominently in the abstract and Section 4; otherwise the paper overclaims robustness.","section":"Appendix F.2, Table 5"},{"comment":"No error bars, repeated-seed runs, or statistical significance measures are reported for any mAP value. The central observation is a relative change in mAP (e.g., 0.4896 vs. 0.6152 for YOLOv5s+AOD-NetX+YOLOv5x), and without variance information it is impossible to assess whether the \"inverse trend\" is a stable effect or within run-to-run noise. Please report at least three independent runs or provide error estimates for the key configurations in Table 1.","section":"Section 4, Table 1; Appendix F"}],"minor_comments":[{"comment":"The phrase \"effective in foggy conditions\" should be qualified as \"effective on in-distribution Foggy Cityscapes,\" given the RTTS results in Table 5.","section":"Abstract and Section 4"},{"comment":"The text says DehazeNet's architecture is \"illustrated in Figure 5,\" but Figure 5 shows dehazing performance on Foggy Cityscapes, not the DehazeNet architecture; the figure reference appears to be incorrect.","section":"Appendix D.3"},{"comment":"The AOD-Net reference appears twice (Li et al., 2017a and 2017b), with the same title and venue; please consolidate.","section":"References"},{"comment":"The title contains an odd hyphenation and spacing artifact (\"CONSE - QUENCES\"); the camera-ready version should fix this and other spacing issues.","section":"Title and front matter"}],"recommendation":"major_revision","confidential_remarks":"The paper is an interesting empirical report, but the central claim currently rests on a confounded comparison and an internal contradiction with Appendix F.1. The requested ablation (AOD-NetX with and without the RoI mask), a correction of the F.1 wording, and error bars would make the contribution much stronger. The OOD results should also be reconciled with the abstract's claims. I do not see grounds for rejection if the authors can address these points within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the arXiv version of From Fog to Failure. The short version: the paper has a real empirical puzzle, but it undercuts itself with an appendix sentence that directly contradicts its main table. That contradiction is the first thing to know.\n\nWhat's actually new: AOD-NetX, a spatial-attention variant of AOD-Net that uses bounding boxes from a lightweight detector to gate dehazing. The authors observe that their AOD-NetX cascade does better on foggy than clear images, while standard YOLO and AOD-Net+YOLO show the usual clear-to-foggy drop. If that observation holds up, it is a useful caution for anyone building cascaded enhancement-detection systems. The related work is reasonably broad, and the discussion of selective preprocessing and haze-gating is sensible.\n\nThe soft spots are serious. Section 4's Table 1 is the entire basis for the headline claim: YOLOv5s+AOD-NetX+YOLOv5x drops clear mAP from 0.5644 to 0.4896 while improving foggy mAP. But Appendix F.1 says, without qualification, that AOD-Net and AOD-NetX 'consistently enhance object detection in both clear and foggy conditions' — and Table 4 in that same appendix reproduces the same clear-image drop. That is an internal contradiction, not a wording nitpick. Second, there are no error bars or repeated runs, and no ablation that removes the RoI attention mask while keeping everything else fixed, so the cause of the clear-image penalty is untested. The 'implicit domain adaptation' explanation is post-hoc. Third, the OOD results in Appendix F.2 show AOD-NetX pipelines underperform the baselines on RTTS and OTS, and Table 3 shows AOD-NetX SSIM collapses to 0.656 on RTTS versus 0.932 for AOD-Net; that sits awkwardly with the 'robust across hazy conditions' framing. On the plus side, the paper does not overclaim in its limitations section: it explicitly calls for haze-gate mechanisms and joint training, citing relevant recent work.\n\nWho gets value from this? Someone working on degradation-aware detection or cascaded enhancement might want this as a data point, but only after the authors fix the contradiction and add the obvious ablations. As it stands, I would not cite it as evidence. A serious referee would give it a major-revision path, not a desk reject, because the empirical direction — that selective preprocessing can hurt clear images — is worth pinning down properly.\n\nRecommendation: send it to review with expectation of heavy revision, but tell the authors that the appendix contradiction and the missing attention ablation are blockers.\n\nBest regards.","headline":"An internally contradictory workshop paper: the headline inverse-trend result is undercut by its own appendix, so the paper needs major revision and targeted ablations before the claim can be trusted.","tokens_in":12331,"tokens_out":2676,"would_cite":false,"duration_ms":25181,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Selective dehazing of detected regions helps object detection in fog but degrades it on clear images.","keywords":["object detection","image dehazing","selective attention","AOD-NetX","foggy conditions","clear image degradation","multi-stage pipeline","Foggy Cityscapes"],"falsifier":"Run AOD-NetX on a clear-image set under three conditions: the original selective attention, an all-ones attention map (global dehazing), and no dehazing. If all-ones keeps clear mAP near baseline while selective attention drops it, the paper's selectivity explanation is supported; if both drop, the penalty is a domain-mismatch artifact. As a second check, retrain AOD-NetX on a mix of clear and foggy images and re-measure Table 1; the inverse trend should disappear if fog overfitting is the cause.","tokens_in":11296,"feed_emoji":"🌫️","tokens_out":9993,"duration_ms":87198,"temperature":0.7,"pith_summary":"The paper tries to establish a counterintuitive result: a selective, vision-inspired dehazing pipeline can help object detection in fog while hurting it in clear weather. It builds a multi-stage framework in which a lightweight detector finds regions of interest, a spatial-attention dehazer (AOD-NetX) enhances only those regions, and a heavier detector makes the final call. On the Foggy Cityscapes data, the YOLOv5s + AOD-NetX + YOLOv5x variant reaches a foggy mAP of 0.6152 against a clear mAP of 0.4896, a 25.68% relative gain under fog; the same inversion appears for the YOLOv8 variant. This matters because any preprocessing step that silently costs accuracy in clear conditions is a liability for systems that must operate in both weathers. The paper attributes the inversion to domain bias in the fog-only-trained dehazer and to overprocessing, and uses it to argue that cascaded enhancement-then-detection pipelines need selective, context-aware triggering.","feed_headline":"Selective dehazing lifts foggy detection, drops clear-image accuracy","feed_subtitle":"A two-stage detector scores higher in fog than in clear weather, so enhancement pipelines need a weather gate.","key_machinery":"The load-bearing object is AOD-NetX, a modification of the AOD-Net dehazing network. Standard AOD-Net takes a hazy input and directly estimates a transmission map $K(x)$ under the atmospheric scattering model; AOD-NetX adds a spatial attention layer built from the bounding boxes that the lightweight detector (YOLOv5s or YOLOv8n) outputs. A sigmoid maps each box's attention weight to $[0,1]$, each region is treated independently, and the attention-weighted transmission map $K'(x)$ is used to dehaze the image. The final heavy detector then runs on the selectively dehazed image. This mechanism is what isolates the paper's claim: the global AOD-Net baseline actually improves clear mAP, so the clear-image penalty appears only when dehazing is attention-gated and trained only on fog.","core_discovery":"The paper's central finding is that AOD-NetX integration reverses the normal clear-to-foggy performance direction. Where baseline YOLOv5x drops from 0.5644 clear mAP to 0.4850 foggy mAP, and global AOD-Net + YOLOv5x drops from 0.6813 to 0.5822, the selective pipeline YOLOv5s + AOD-NetX + YOLOv5x scores 0.4896 clear and 0.6152 foggy. The same shape appears with YOLOv8n + AOD-NetX + YOLOv8x: 0.5150 clear versus 0.6114 foggy. The paper reads these numbers as an implicit domain-adaptation effect: AOD-NetX's features are optimized for haze removal and therefore distort clear images, while helping foggy ones. It also confirms that the benefit is dataset-specific, because on the RESIDE-$\\beta$ OTS and RTTS out-of-distribution sets the AOD-NetX variants fall below their baselines.","pith_inferences":["The paper does not fully isolate the attention mask from the training-domain mismatch; if a fog-only-trained dehazer distorts clear pixels regardless of where it is applied, the clear-image penalty would not be evidence against selective enhancement specifically.","A natural experiment the paper leaves implicit is to set AOD-NetX's attention map to all ones on clear images and compare with the selective map: equal degradation would point to domain mismatch, while unequal degradation would point to the mask.","The inverse trend could be partly a paired-data artifact, because the clear images in Foggy Cityscapes are the same scenes used to make the foggy images; testing on an independent clear benchmark would settle how much of the clear penalty is real."],"forward_implications":["Deploying a fog-optimized selective dehazer in an always-on system would trade clear-weather accuracy for foggy-weather accuracy, so a domain or haze-level gate is needed before enhancement is applied.","AOD-NetX's foggy gains do not transfer across fog distributions: on RESIDE-$\\beta$ OTS and RTTS, the AOD-NetX variants score below the YOLOv5x and YOLOv8x baselines, so the phenomenon is not a general fog benefit.","The contrast with global AOD-Net indicates that the clear-image degradation is tied to the selective attention mechanism and its fog-only training, not to dehazing as such.","Training or fine-tuning the whole enhancement-plus-detection pipeline jointly, or adding clear images to the dehazer's training set, becomes a necessary design step rather than an optional refinement."],"supporting_citations":[{"why":"Supplies AOD-Net, the base dehazing network whose transmission map AOD-NetX modifies; the global AOD-Net baseline in Table 1 is the key contrast for isolating the clear-image penalty.","marker":"Li et al., 2017b"},{"why":"Provides Foggy Cityscapes, the synthetic fog dataset on which the inverted clear-to-foggy mAP trend is measured.","marker":"Sakaridis et al., 2018"},{"why":"Provides RESIDE-$\\beta$, OTS, and RTTS, the out-of-distribution benchmarks showing that AOD-NetX's foggy gains do not transfer.","marker":"Li et al., 2019"},{"why":"DehazeNet is one of the dehazing baselines compared before the paper selects AOD-Net as the base for AOD-NetX.","marker":"Cai et al., 2016"},{"why":"Dehaze-UNet is the other dehazing baseline compared, establishing that the choice of AOD-Net is empirical rather than assumed.","marker":"Zhou et al., 2024"}],"fun_headline_variants":["Dehazing pipeline flips detection: better in fog, worse in sun","Clear images hurt by dehazing boost, fog benefits","Hybrid detector gains in fog, loses in clarity","Why adding dehazing to detection backfires on clear scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the spatial attention map built from the lightweight detector's bounding boxes is a faithful implementation of selective region enhancement, so the clear-image penalty is caused by that selectivity combined with fog-only training, not by applying a fog-trained network to clear inputs.","fun_headline_variants_meta":{"raw":{"variants":["Dehazing pipeline flips detection: better in fog, worse in sun","Clear images hurt by dehazing boost, fog benefits","Hybrid detector gains in fog, loses in clarity","Why adding dehazing to detection backfires on clear scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1205,"prompt_tokens":890,"completion_tokens":315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":246}},"tokens_in":506,"tokens_out":315,"duration_ms":3719,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:37:58.852123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AOD-NetX on a clear-image set under three conditions: the original selective attention, an all-ones attention map (global dehazing), and no dehazing. If all-ones keeps clear mAP near baseline while selective attention drops it, the paper's selectivity explanation is supported; if both drop, the penalty is a domain-mismatch artifact. As a second check, retrain AOD-NetX on a mix of clear and foggy images and re-measure Table 1; the inverse trend should disappear if fog overfitting is the cause.","supporting_citations":[],"review_version":1}