{"id":"5afe9ce3-ffe8-408a-838a-079104b5b254","arxiv_id":"2508.14597","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A Swin-transformer smoke detector is trained on segmentation masks produced by the authors' own optical flow and Gaussian mixture pipeline, with reported accuracy and uncertainty claims lacking independent human-annotated evaluation.","lead":"This paper builds a smoke-detecting AI system in three steps: it computes motion in video with its own optical flow model, turns that motion into binary smoke labels, and trains a transformer to reproduce those labels with confidence scores. The system matters because early smoke detection is safety-critical, but its accuracy numbers measure agreement with labels its own software created, so independent validation is essential.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation loop: smoke masks are generated by the paper's own GMM pipeline and used as ground truth, so reported accuracy/calibration may reflect front-end bias, not true smoke detection.","rationale":"The reader's weakest assumption—that the GMM-generated masks are treated as unbiased ground truth—is exactly the load-bearing concern. The manuscript's explicit wording makes this a documented circularity rather than speculation: the dataset is constructed from the authors' own optical flow and GMM binarization, and the same pipeline supplies the supervision and the evaluation targets. The strongest claim about reliability and calibration cannot survive if those masks are biased. I see no other concern of comparable weight. The optical flow component does receive external validation on Middlebury and Sintel, which is a point in the paper's favor, but it does not validate the smoke-specific labeling. The main body is largely unreadable mojibake in the supplied version, so the assessment relies on the abstract, Table 1, and the readable page fragments. If the full text contains a separate human-annotated evaluation, the verdict should be revisited; the proposed concrete test would settle whether the concern actually lands. Since the reader's verdict is REJECT with low confidence and our analysis supports that rejection, no change to the verdict is needed.","tokens_in":5152,"tokens_out":2492,"duration_ms":32016,"concrete_test":"Select a random sample of 200 frames from the smoke dataset. Have at least three independent annotators with fire-safety domain expertise create pixel-level smoke masks (majority vote as consensus). Compute IoU/Dice between the GMM-generated masks and the human consensus. Then evaluate the trained TP-UAST on a held-out subset of the human-labeled frames (e.g., 80/20 split) and compare segmentation mIoU and reliability diagrams against the reported numbers. If human-GMM IoU < 0.7, or if TP-UAST's accuracy on the human-labeled test set is materially lower (e.g., >5–10 points mIoU) than reported, the self-generated labels are the source of the perceived performance and the generalization/calibration claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—superior generalization, robustness, and well-calibrated uncertainty—rests on a self-referential evaluation. The abstract states that color-encoded optical flow maps are fused with appearance cues via GMM 'to generate binary segmentation masks of the smoke regions.' Page 13 confirms: 'This binary mask is derived from the estimated color maps using the GMM technique.' These masks are used both to train the transformer and to measure its segmentation accuracy and uncertainty calibration. Consequently, the reported metrics measure how well TP-UAST reproduces the front-end's own labeling, not how well it detects actual smoke. The only external validation, Table 1 on Middlebury and Sintel, tests the optical flow model, not the smoke masks. If the GMM systematically mislabels non-smoke motion (vehicles, trees, pedestrians) as smoke, the high scores and 'well-calibrated' uncertainty are artifacts. This is an internal-validity problem, not a disagreement with external consensus, and it directly undermines the reliability claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TP-UAST, a Two-Phase Uncertainty-Aware Shifted Windows Transformer for smoke segmentation from monocular video. The pipeline first estimates optical flow with a fractional-order variational model, color-encodes the flow, fuses it with appearance cues via a Gaussian Mixture Model (GMM) to produce binary smoke masks, and then trains a Swin-Transformer-based segmentation network with a two-phase loss: first for segmentation accuracy, then for aleatoric and epistemic uncertainty estimation. The authors claim superior generalization, robustness, and well-calibrated uncertainty relative to state-of-the-art smoke detectors. The only external validation reported is for the optical-flow component on Middlebury and Sintel; the smoke segmentation and calibration claims are evaluated using masks generated by the same GMM pipeline that is part of the proposed system.","tokens_in":5362,"tokens_out":3832,"duration_ms":47747,"significance":"If the central evaluation were independent, the paper would contribute a useful integration of optical-flow-guided pseudo-labeling, transformer-based segmentation, and joint aleatoric/epistemic uncertainty estimation for early fire detection. The optical flow module is benchmarked against external datasets (Table 1), and the two-phase uncertainty training is a sensible architectural choice. However, the significance of the paper hinges on the validity of the smoke masks used as ground truth. Because those masks are generated by the paper's own front-end, the reported segmentation accuracy and calibration metrics currently measure how well the transformer reproduces the GMM labels, not how well it detects actual smoke. Independent validation is therefore essential before the claimed reliability can be accepted.","major_comments":[{"comment":"The smoke masks used to train and evaluate TP-UAST are produced by the same pipeline that is being proposed. The abstract states that 'color-encoded optical flow maps are fused with appearance cues via a Gaussian Mixture Model to generate binary segmentation masks of the smoke regions,' and p. 13 confirms: 'This binary mask is derived from the estimated color maps using the GMM technique.' These masks serve as ground truth for both training and scoring the transformer, making the reported accuracy and calibration self-referential. If the front-end systematically labels non-smoke motion (vehicles, trees, pedestrians) as smoke, the high scores and 'well-calibrated' uncertainty are artifacts of the front-end bias rather than evidence of reliable smoke detection. Please evaluate on human-annotated smoke segmentation datasets, or at least on a manually validated subset, and report per-image a","section":"Abstract and p. 13, 'Fourth experiment'"},{"comment":"The only external benchmarks reported (Middlebury, Sintel) validate the optical flow model, not the smoke segmentation or the uncertainty calibration. Table 1 therefore does not support the abstract's claim of 'superior generalization and robustness' for the detection system as a whole. A quantitative comparison against established smoke/fire detection benchmarks or human-annotated smoke masks is needed to substantiate the central claim. Without such independent evaluation, the reported detection performance cannot be distinguished from the behavior of the GMM labeling step.","section":"Table 1"},{"comment":"The claim that TP-UAST's 'uncertainty estimates are well-calibrated, informative of error, and sensitive to class-dependent difficulty' is made using the same self-generated labels discussed above. In addition, no quantitative calibration metric (e.g., Expected Calibration Error, Brier score) is reported; the evidence consists of reliability diagrams and histograms. Please report standard calibration metrics and compare them with baseline uncertainty methods, using independent ground-truth labels rather than the pipeline's own masks.","section":"pp. 15–17, Figs. 11–12"}],"minor_comments":[{"comment":"The 'novel smoke segmentation dataset' is mentioned in the abstract but not described in sufficient detail. Please include image/video counts, sources, resolution, split information, and the rationale for the GMM component count and binarization threshold. Without this information, the experiments are not reproducible and the sensitivity of the generated labels to these choices cannot be assessed.","section":"Dataset description"},{"comment":"Qualitative comparisons on smoke and non-smoke data are presented, but no quantitative false-positive / false-negative rates are given for the non-smoke cases. A quantitative analysis of false positives on non-smoke videos would strengthen the reliability claim.","section":"Figures 7–8"},{"comment":"Several passages of the manuscript text are garbled or contain unreadable characters, making it difficult to verify the mathematical formulation in Sections 2–3. A clean, correctly rendered version should be provided.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The self-referential evaluation loop is the central issue: the smoke masks used as ground truth are produced by the paper's own GMM front-end, and the only external validation concerns the optical flow module. This is a correctable evaluation gap, but it requires substantial additional experiments on independent smoke annotations or a recognized benchmark. I therefore recommend major revision rather than reject, because the central idea is defensible if independent validation confirms the labels."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real product is an assembly: a fractional-order variational optical flow model (their own prior work, relabeled FCDLe-FOV), a GMM that binarizes color-encoded flow into smoke masks, and a Swin Transformer with a two-phase uncertainty head. That assembly is new in a weak sense, and the optical flow component gets some legitimate external validation on Middlebury and Sintel in Table 1. The two-phase training idea is also standard but sensible. If the paper were only about optical flow, I'd say it's a reasonable incremental contribution.\n\nThe problem is the central claim: reliable smoke segmentation with well-calibrated uncertainty. The evaluation for that claim is circular. The abstract and page 13 both say the binary masks used to train and score the transformer are generated by the paper's own GMM from its own optical flow color maps. So the reported accuracy, confusion matrix, and reliability diagram measure how well the transformer reproduces the front-end's biases, not how well it detects actual smoke. If the GMM mislabels moving vehicles or trees as smoke, the metrics will look great and the uncertainty will look calibrated to that error, not to real smoke. There is no independent human-labeled smoke dataset, no code, no data visible. The only external benchmarks stop at the flow module.\n\nThe mojibake in the body text makes this worse, because I can't check whether there is a hidden evaluation against real smoke annotations. Based on what is readable, the self-referential loop is load-bearing.\n\nThe authors clearly know the technical pieces, and the flow model is not phoned in. But the paper's selling point—superior generalization and reliable early fire detection—is unsupported as written. A serious referee could not salvage this without a fundamental redesign of the evaluation: separate human-annotated masks for training and testing, or at minimum an independent test set with real smoke and non-smoke motion.\n\nFor peer review, I would not send this out for full review. It's a desk reject in its current form. If the authors add a genuine independent evaluation, revisit it. For a reading group, it's a useful case study in how easy it is to build a closed evaluation loop with a self-generated dataset, but not for the technical method.","headline":"The smoke-detection accuracy and calibration claims are built on a self-referential evaluation loop, so the headline numbers are not trustworthy; the optical flow module is externally validated, but that is not the paper's central claim.","tokens_in":5886,"tokens_out":1503,"would_cite":false,"duration_ms":22234,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single-camera pipeline claims to segment smoke from video and flag, with a calibrated confidence score, when its own prediction may be wrong.","keywords":["smoke detection","optical flow","level-set variational model","Swin transformer","uncertainty estimation","Gaussian mixture segmentation","early fire detection","monocular video segmentation"],"falsifier":"Run the published flow-to-GMM-to-transformer pipeline on a human-annotated test set of real fire and smoke videos and compare per-pixel agreement against the manual masks; if agreement falls well short of the reported numbers, or if the model outputs high-confidence smoke masks on smoke-free clips containing moving vehicles or pedestrians, the synthetic-label premise is broken. A simpler version: a video of a car crossing a smoke-free scene should produce no smoke mask, so any substantial mask from non-smoke motion directly indicts the GMM stage.","tokens_in":5021,"feed_emoji":"🔥","tokens_out":6568,"duration_ms":71057,"temperature":0.7,"pith_summary":"This paper claims that smoke can be detected reliably from ordinary single-camera video, without multi-sensor hardware, by fusing motion and appearance cues and letting a transformer both segment the smoke and measure its own uncertainty. The training labels are generated automatically: a fractional-order variational optical flow model color-encodes motion while preserving smoke-boundary discontinuities, and a Gaussian Mixture Model binarizes those maps into smoke masks. A two-phase shifted-windows transformer first learns accurate segmentation, then learns to output aleatoric and epistemic uncertainty, with the claim that the resulting confidence is well-calibrated, tracks error, and reflects per-class difficulty. If true, surveillance and industrial safety systems could issue early fire warnings with an explicit reliability signal attached to every prediction.","feed_headline":"Transformer segments smoke and scores its own certainty","feed_subtitle":"Optical flow builds the labels; a two-phase Swin transformer adds calibrated uncertainty for early fire alarms.","key_machinery":"The load-bearing object is the two-phase uncertainty-aware shifted-windows transformer (TP-UAST), built on a Swin Transformer backbone with a multi-scale uncertainty head. Its labels come from two coupled mechanisms: an optical-flow stage (FCDLe-FOV, a four-color-theorem-inspired dual-phase level-set fractional-order variational model) that color-encodes motion while preserving discontinuities, and a Gaussian Mixture Model that fuses the flow color maps with appearance cues into binary smoke masks. The two-phase training regimen is the third mechanism: phase one fixes segmentation accuracy, and phase two fits aleatoric and epistemic uncertainty so the model outputs a plausibility score along","core_discovery":"The central claim is that a Two-Phase Uncertainty-Aware Shifted Windows Transformer (TP-UAST), trained on masks produced by its own optical-flow-plus-GMM pipeline, both segments smoke and estimates how much its own prediction should be trusted. The optical flow stage — a four-color-theorem-inspired dual-phase level-set fractional-order variational model — encodes motion into color maps while preserving the sharp discontinuities at smoke boundaries; a Gaussian Mixture Model then fuses these motion maps with appearance cues into binary smoke masks. The first training phase optimizes segmentation accuracy; the second teaches the model to emit a plausibility confidence per prediction by jointly","pith_inferences":["The paper leaves open whether its calibration generalizes: because both training and evaluation use the same synthetic masks, 'well-calibrated' may mean consistent with the pipeline's own labels rather than with human judgment of smoke; a human-annotated held-out set is the decisive check.","The identical flow-to-GMM labeling trick should transfer to other semi-transparent moving phenomena such as steam, fog, or dust, where manual segmentation is equally hard — with the same circular-label caveat applying.","Because the uncertainty head separates aleatoric from epistemic components, the framework is a natural fit for active learning: the model could request human labels for the frames where epistemic uncertainty is highest.","Deployment risk concentrates in the motion prior: if the GMM stage labels any moving object (vehicles, trees, pedestrians) as smoke, the reported accuracy measures agreement with that prior, not true smoke detection, so domain shift could silently break the system."],"forward_implications":["Early-warning alarm systems could run on ordinary surveillance footage from a single camera, with no multi-sensor array, because smoke precursors are detected from motion-encoded monocular video.","Every prediction carries a plausibility confidence, so alarm logic can defer to human review precisely when the model reports high uncertainty rather than firing on all detections.","The smoke-segmentation dataset is produced automatically from flow and color cues, removing the need for costly manual pixel-level annotation of smoke.","The flow model's discontinuity preservation is claimed to yield sharper smoke boundaries than prior variational and learned optical-flow baselines on Middlebury and Sintel benchmarks.","Separating accuracy from confidence lets the calibration phase be re-run or extended without retraining the segmentation head, since the two tasks are optimized in distinct phases."],"supporting_citations":[{"why":"Supplies the level-set fractional-order variational motion-estimation model that FCDLe-FOV extends and is compared against.","marker":"[5]"},{"why":"Supplies the shifted-windows transformer backbone that TP-UAST augments with its multi-scale uncertainty head.","marker":"[30]"},{"why":"Supplies the Gaussian Mixture Model technique used to binarize the optical flow color maps into smoke-segmentation labels.","marker":"[7]"},{"why":"Provides the uncertainty–error correspondence methodology behind the paper's calibration analysis.","marker":"[8]"},{"why":"Prior CNN-based fire prediction using fractional-order optical flow whose smoke-mask output is the comparison baseline for the GMM masks.","marker":"[9]"},{"why":"Field-segmentation variational optical flow baseline (FS-FOV) that Table 1 shows FCDLe-FOV outperforming.","marker":"[34]"},{"why":"Modified variational functional baseline (HS-NE) used for optical flow accuracy comparison on Middlebury and Sintel.","marker":"[21]"},{"why":"Non-local weighted regularization baseline (NLW) used for optical flow accuracy comparison on Middlebury and Sintel.","marker":"[22]"}],"fun_headline_variants":["Smoke detection nets learn to doubt their own calls","Two-phase transformer segments smoke, gauges its confidence","Optical flow builds labels; Swin transformer adds self-doubt","Smoke segmenter trained on its own masks, plus uncertainty","Swin transformer both finds smoke and rates its own certainty"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The training labels and the evaluation ground truth are both produced by the paper's own pipeline — a Gaussian Mixture Model binarizing its own optical-flow color maps — so the reported accuracy and confidence calibration presuppose that these synthetic masks faithfully identify smoke and nothing else.","fun_headline_variants_meta":{"raw":{"variants":["Smoke detection nets learn to doubt their own calls","Two-phase transformer segments smoke, gauges its confidence","Optical flow builds labels; Swin transformer adds self-doubt","Smoke segmenter trained on its own masks, plus uncertainty","Swin transformer both finds smoke and rates its own certainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1837,"prompt_tokens":776,"completion_tokens":1061,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":979}},"tokens_in":520,"tokens_out":1061,"duration_ms":8797,"temperature":1.0,"reasoning_tokens":979,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:24:17.719394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published flow-to-GMM-to-transformer pipeline on a human-annotated test set of real fire and smoke videos and compare per-pixel agreement against the manual masks; if agreement falls well short of the reported numbers, or if the model outputs high-confidence smoke masks on smoke-free clips containing moving vehicles or pedestrians, the synthetic-label premise is broken. A simpler version: a video of a car crossing a smoke-free scene should produce no smoke mask, so any substantial mask from non-smoke motion directly indicts the GMM stage.","supporting_citations":[{"cited_title":"A level set based fractional order variational model for motion estimation in application oriented spectrum","cited_arxiv_id":null,"evidence_quote":"Supplies the level-set fractional-order variational motion-estimation model that FCDLe-FOV extends and is compared against."},{"cited_title":"Image segmentation using gaussian mixture model","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian Mixture Model technique used to binarize the optical flow color maps into smoke-segmentation labels."},{"cited_title":"Improving Uncertainty-Error Correspondence in Deep Bayesian Medical Image Segmentation","cited_arxiv_id":"2409.03470","evidence_quote":"Provides the uncertainty–error correspondence methodology behind the paper's calibration analysis."},{"cited_title":"Cnn-based fire prediction using fractional order optical flow and smoke features","cited_arxiv_id":null,"evidence_quote":"Prior CNN-based fire prediction using fractional-order optical flow whose smoke-mask output is the comparison baseline for the GMM masks."},{"cited_title":"A field-segmentation-based variational optical flow method for piv measurements of nonuniform flows","cited_arxiv_id":null,"evidence_quote":"Field-segmentation variational optical flow baseline (FS-FOV) that Table 1 shows FCDLe-FOV outperforming."},{"cited_title":"A modified variational functional for estimating dense and discontinuity preserving optical flow in various spectrum","cited_arxiv_id":null,"evidence_quote":"Modified variational functional baseline (HS-NE) used for optical flow accuracy comparison on Middlebury and Sintel."},{"cited_title":"Non-local weighted regularization for optical flow estimation.Optik, 208:164069, 2020","cited_arxiv_id":null,"evidence_quote":"Non-local weighted regularization baseline (NLW) used for optical flow accuracy comparison on Middlebury and Sintel."}],"review_version":1}