{"id":"586e0d02-10e2-49f8-92bb-99a709a99938","arxiv_id":"2507.21959","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A teacher-student CNN/ViT framework with feature-level consistency raises weakly supervised smoke-segmentation seed mIoU from 33.56 to 47.37 and to 52.93 with post-processing on a custom IJmond dataset.","lead":"This paper trains a smoke-detection model on images with only class labels, not pixel masks, by aligning the internal features of a CNN and a Vision Transformer so their individual biases cancel out. It reports large gains on a private industrial smoke dataset, but the evaluation relies on test-set-tuned parameters and one dataset, so the headline numbers should be read with caution.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The co-occurrence-suppression claim is not actually measured: test mIoU is computed only on smoke-containing patches, so chimney-without-smoke patches, the direct probe of spurious chimney activation, are excluded from evaluation.","rationale":"The reader identified the same weakest assumption: the evaluation protocol restricts scoring to smoke-containing patches, and hyperparameters are selected on the test set. I agree this is the most load-bearing issue, because the paper's novelty claim is specifically about mitigating co-occurrence bias, not about improving CAM quality in general. The internal ablations (Table 4.7, 4.8) do show that the teacher-student cosine global variant outperforms other transfer strategies and the PCM baseline, which is real evidence that the framework helps on the chosen metric. However, mIoU on smoke-containing patches is a composite measure; it does not by itself demonstrate that chimney activations are suppressed in the absence of smoke. The paper's own Section 6.2 defers other datasets and evaluations to future work, and the dataset description in Section 4.1 explicitly mentions non-smoke scenes, so a chimney-only test is feasible. The concern is not that the method is wrong, but that the current experiments do not test the specific phenomenon the title and central claim assert. This warrants a conditional verdict: the method is promising and internally consistent, but the co-occurrence-suppression claim needs a dedicated evaluation before it can be accepted.","tokens_in":20214,"tokens_out":2500,"duration_ms":33177,"concrete_test":"Construct a chimney-only evaluation set from the IJmond non-smoke scenes (chimney visible, no smoke). Run the trained teacher-student model and the PCM/ResNet50 baseline on these patches and report smoke-class false positive rate and mIoU with all pixels treated as background. If the cross-architecture model does not reduce chimney false positives relative to PCM, the claim that consistency regularization suppresses co-occurrence bias is not supported by the current evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that cross-architecture feature consistency suppresses co-occurrence bias (chimney-smoke spurious correlation) without external supervision. The supporting evidence is seed pseudo-mask mIoU (47.37 vs 33.56 for PCM), but Section 4.1 states: 'Only patches contains visible smoke will be used for evaluation.' This means chimney-without-smoke patches are never scored. The strongest form of the claimed bias is that the model activates chimney regions even when no smoke is present, because chimneys and smoke always co-occur in the training distribution. On the current test set, such spurious activations cannot be detected: if the model maps the whole chimney to smoke, it is penalized only when that chimney happens to overlap a smoke-containing patch, and even then the mIoU conflates foreground coverage, boundary quality, and co-occurrence suppression into one number. A model that improves CAM coverage or boundary sharpness without reducing chimney false positives would still show higher mIoU, so the reported gains do not isolate the mechanism the paper claims to introduce. The qualitative figures (Figure 4.1, 4.2) show examples, but no quantitative false-positive or chimney-only evaluation is provided. The additional lack of a validation split (lambda in Table 4.10, CRF scaling in Table 4.11c, SAM points and fusion in Tables 4.11a/b selected on the test set) further inflates the headline numbers, but the evaluation-gap concern is more directly load-bearing: even if the numbers were honest, they would not establish co-occurrence suppression.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Zhang and Hsu propose a weakly supervised semantic segmentation (WSSS) method for industrial smoke that uses only image-level labels. The core idea is a teacher-student framework in which a pretrained ResNet teacher guides a ViT student through a feature-level consistency loss (global cosine similarity after a linear projection), motivated by the observation that CNNs and ViTs produce complementary CAMs. They also evaluate a co-training variant and a set of post-processing steps (CRF, SAM-enhanced, CAM fusion, multi-scale testing). On a custom IJmond smoke dataset, the method reports seed pseudo-mask mIoU of 47.37 versus 33.56 for PCM, and 52.93 after post-processing, and claims to be the first to suppress co-occurrence bias via cross-architecture consistency without external supervision.","tokens_in":20394,"tokens_out":8243,"duration_ms":95593,"significance":"The central idea is worth taking seriously: using two heterogeneous architectures as mutual regularizers is a plausible way to reduce spurious correlations in WSSS, and the ablations are more systematic than is typical for this area. The paper provides code, compares teacher-student versus co-training, three alignment levels, several similarity metrics, feature projection, and a lambda sweep; the consistency loss improves mIoU over a wide lambda range (38.55 at 0.3 to 47.37 at 1.0 versus the 33.56 baseline), which is genuine evidence that the regularizer helps in the current evaluation. However, the evaluation protocol cannot currently support the headline claim of co-occurrence suppression: the test set is restricted to patches containing visible smoke, and hyperparameters are selected on the same test set that produces the final numbers. These issues are fixable, but until fixed the claimed significance is not established.","major_comments":[{"comment":"The evaluation protocol cannot support the paper's central claim that cross-architecture consistency suppresses co-occurrence bias. Section 4.1 states that 'Only patches contains visible smoke will be used for evaluation.' Because the co-occurrence problem is precisely that chimney regions activate even when no smoke is present, excluding chimney-without-smoke patches removes the direct probe of the phenomenon. The reported mIoU gains (47.37 versus 33.56 for PCM) could be produced by improved foreground coverage or boundary sharpness without any reduction in spurious chimney activation. Please evaluate on all test patches, including smoke-free chimney patches, and report a false-positive metric for chimney regions, or explicitly report per-pixel precision on non-smoke areas in addition to smoke-class IoU.","section":"Section 4.1; Tables 4.1 and 4.3"},{"comment":"No validation split is described, and all tunable hyperparameters are selected by maximizing mIoU on the same 900-image test set used to report the final numbers. Specifically, lambda in Table 4.10 is set to 1.0 because it gives the maximum mIoU there; SAM points per side (Table 4.11a), SAM fusion strategy (Table 4.11b), and CRF scaling (Table 4.11c) are likewise chosen by test-set performance. The 'Optimal threshold' rows in Table 4.4 (53.92 and 57.15) are ground-truth oracle numbers and should not be presented alongside achievable methods. Please introduce a held-out validation split or nested cross-validation, select hyperparameters on it, and report final test numbers only once; remove or clearly label oracle rows.","section":"Section 4.4, Tables 4.10 and 4.11; Section 4.1"}],"minor_comments":[{"comment":"Please fix the sentence 'Only patches contains visible smoke will be used for evaluation' and specify how many patches contain smoke, how many do not, and whether non-smoke patches are entirely absent from the test evaluation.","section":"Section 4.1"},{"comment":"The row 'image-level + limited pixel-level T1 ViT-B✗' has no reported mIoU and an unexplained symbol; either report the value or remove the row.","section":"Table 4.1"},{"comment":"The 'Seed' column is constant 33.56 in all rows; clarify that this is the PCM baseline seed and show the seed value for each variant, since seed quality depends on the training paradigm and projection choice.","section":"Table 4.8"},{"comment":"The claim of 'no external supervision' should be qualified: the backbones are initialized with ImageNet classifier weights, which is external pretraining even if no additional task-specific labels or priors are used.","section":"Section 4.2"},{"comment":"The text 'citeshu2021channel proposed' contains an unresolved citation key; please replace it with the proper reference.","section":"Section 2.2"},{"comment":"The limitations paragraph lists underexplored research directions but does not mention the evaluation protocol limitations, specifically the smoke-only patch evaluation and the absence of a validation split; these should be acknowledged.","section":"Section 6.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the ablations are systematic, but the evaluation protocol is the main risk. The test-set-only evaluation and test-set hyperparameter selection are fixable with a proper validation split and a direct chimney false-positive metric; I would encourage the editor to request those changes rather than reject. The paper also needs editorial tightening: broken citations, incomplete table rows, and the misleading 'Optimal threshold' rows suggest it is not yet polished for journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: the paper has a plausible mechanism and genuine systematic ablations, but the headline claim about suppressing co-occurrence is not actually measured by its test protocol. The stress-test is right: Section 4.1 says only patches with visible smoke are evaluated, so the cleanest probe for chimney false positives—chimney without smoke—is excluded. The mIoU gains could come from better foreground coverage or boundary sharpness, not from reducing co-occurrence activation. Even if every number is honest, the experiments do not isolate the claimed mechanism.\n\nWhat is genuinely new: teacher-student ResNet-to-ViT feature consistency with a 1x1-conv projection and global cosine loss, targeted at co-occurrence in WSSS. The ablation set is thorough—paradigm, metric, alignment level, projection, loss weight—and the consistency loss improves seed mIoU across a wide lambda range (38.55 to 47.37 vs 33.56 baseline). That is real evidence the regularizer does something. The teacher-student vs co-training comparison is informative: the frozen pretrained teacher wins, which is a useful data point.\n\nSoft spots, in order of severity. First, the evaluation-gap issue above. Second, no validation split: lambda, CRF scaling, SAM points/fusion are all chosen on the same 900-image test set that produces the headline numbers. Third, no error bars or multiple seeds. Fourth, missing comparisons to the closest cross-architecture baselines the paper itself cites ([40], [41], [34]). The \"no external supervision\" claim is overstated given ImageNet pretraining, frozen SAM, and out-of-distribution training images. These are constraints, not refutations: the method is demonstrated on a single custom dataset, not confirmed.\n\nThe paper reads like a master's thesis, which it likely is: honest about limitations (Section 6.2), unpolished in places, but clear about what was tried. For a reader in WSSS or applied smoke monitoring, it is a useful case study in how test-set selection can inflate gains. I would send it to a serious referee, asking for a validation split, a chimney-only evaluation, and the missing baselines. Releasing the dataset and code would make it much easier to assess.","headline":"Plausible mechanism and systematic ablations, but the test protocol never measures the claimed co-occurrence suppression, and test-set hyperparameter selection inflates the headline numbers.","tokens_in":21101,"tokens_out":3044,"would_cite":false,"duration_ms":32567,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Forcing a ResNet teacher and a Vision Transformer student to align their internal feature representations suppresses co-occurrence bias in weakly supervised smoke segmentation, with no external supervision or prior knowledge.","keywords":["weakly supervised semantic segmentation","spurious correlations","co-occurrence bias","cross-architecture consistency","teacher-student knowledge transfer","class activation maps","vision transformer","industrial smoke segmentation"],"falsifier":"Score the trained student model on chimney-only patches — chimney present, no visible smoke — and measure the smoke-class CAM activation: if a meaningful fraction of such patches light up, the spurious correlation is not suppressed and the reported mIoU, computed only on patches with visible smoke, would be hiding it. A companion check is to re-tune $\\lambda$, the CRF scaling, the SAM points-per-side, and the SAM fusion strategy on a held-out split of the same dataset; if the 47.37 and 52.93 figures shrink substantially, a portion of the gain reflects selection rather than the consistency mechanism.","tokens_in":19842,"feed_emoji":"🌫️","tokens_out":28998,"duration_ms":257577,"temperature":0.7,"pith_summary":"The paper tries to establish that spurious correlations — a weakly supervised model's habit of treating a chimney as part of the smoke that always appears with it — can be suppressed with no extra labels, no external priors, and no human knowledge. In weakly supervised semantic segmentation (WSSS), only image-level labels are available, so pixel masks for training a segmentation model are derived from class activation maps (CAMs), the heatmaps of which regions drive the classifier's decision; these maps are prone to firing on co-occurring context like chimneys. The authors' remedy is a teacher-student setup in which a pretrained ResNet teacher guides a Vision Transformer student via a consistency loss that aligns their internal feature representations, exploiting the observation that the two architectures make different mistakes: the CNN localizes smoke but blurs its boundaries, while the ViT draws sharp boundaries but also activates co-occurring context. On a custom industrial-smoke dataset, the method raises the seed-mask mean intersection-over-union (mIoU), a standard overlap score, from 33.56 for the strongest comparison baseline (PCM, a pixel-correlation method) to 47.37, and to 52.93 after adding post-processing that combines the Segment Anything Model (SAM) with a conditional random field (CRF). The broader claim is that the agreement enforced between heterogeneous architectures acts as an internal supervisor, so a pair of biased classifiers can correct each other's bias without any outside knowledge.","feed_headline":"Aligned ResNet and ViT lift smoke-spotting from 33.6 to 47.4","feed_subtitle":"A consistency loss suppresses chimney-smoke confusion using only image-level labels — no extra supervision.","key_machinery":"The load-bearing mechanism is a knowledge transfer consistency loss inside a teacher-student framework. A frozen, pretrained ResNet-50 acts as teacher and a trainable ViT-B as student; each network's feature map first passes through a learnable linear projection (a 1×1 convolution) into a shared two-channel space suited to the binary smoke/background task, and after global average pooling the two resulting vectors are aligned by cosine similarity, $S_{global} = \\frac{f^{(v)} \\cdot f^{(r)}}{\\| f^{(v)} \\| \\| f^{(r)} \\|}$. The training objective is $\\mathcal{L} = \\mathcal{L}_{cls} + \\lambda \\mathcal{L}_{global}$ with $\\mathcal{L}_{global} = 1 - S_{global}$, where $\\lambda$ is a tuned weight. This loss forces the ViT to behave internally more like the CNN, transferring the CNN's ability to ignore co-occurring context while keeping the ViT's sharper boundaries; ablations show the teacher-student configuration with global alignment and cosine similarity outperforms co-training, spatial or channel-wise alignment, and $\\ell^1$ or $\\ell^2$ distances.","core_discovery":"The paper's central claim is that feature-level consistency between two architectures, enforced during training, suppresses co-occurrence bias in weakly supervised segmentation without any external supervision. Concretely, a fixed pretrained ResNet-50 teacher and a trainable ViT-B student are trained together with a binary cross-entropy classification loss plus a knowledge transfer loss that maximizes the cosine similarity of their global average-pooled feature vectors, after each network's features pass through a learnable linear projection into a shared space. The resulting CAMs activate foreground smoke rather than the co-occurring chimney, raising seed pseudo-mask mIoU to 47.37 on the custom smoke dataset, compared with 33.56 for PCM, 26.10 for a plain ResNet50, and 13.18 for a ViT-S trained alone; after CRF and SAM-enhanced post-processing, the pseudo masks reach 52.93. The authors state that they are the first to report that cross-architecture consistency alone can suppress co-occurrence bias during representation learning, and they read the result as evidence that heterogeneous inductive biases supply complementary supervision that needs no external source.","pith_inferences":["The recipe is architecture-agnostic: any two networks with different inductive biases and different failure modes could be aligned the same way, so the method is directly testable in other co-occurrence-heavy domains, such as medical imaging where a lesion always appears with an organ, or satellite imagery where a target always appears with fixed infrastructure.","The cleanest test the paper does not run is to score chimney-only patches — chimney present, no visible smoke — where a model that truly suppresses co-occurrence bias should produce near-zero smoke activation; that false-positive rate would settle the claim more directly than mIoU on the smoke-only test patches.","Two caveats bound the reported numbers: the loss weight, CRF scaling, SAM points per side, and SAM fusion strategy were all selected by maximizing mIoU on the same test set that produced the gains, and the authors' own ViT-B baseline trained with image-level labels on a portion of the test set reaches 47.99 mIoU, within 0.62 of the proposed method's 47.37.","The frozen, pretrained teacher sets a ceiling on the student: because the ViT is pulled toward the ResNet's representation, the method inherits the teacher's blind spots, so a stronger teacher or one that keeps updating is the most direct route to further seed improvement."],"forward_implications":["Seed pseudo-mask quality for weakly supervised smoke segmentation rises to 47.37 mIoU with only image-level labels, up from 33.56 for PCM, 26.10 for a plain ResNet50, and 13.18 for a ViT-S trained alone.","Appending CRF and SAM-enhanced post-processing raises the pseudo masks to 52.93 mIoU, within reach of fully supervised references (54.68 for a fine-tuned SAM, 68.27 for a transformer segmentation model).","Because the co-occurrence bias is removed during representation learning rather than at refinement time, the approach targets what the paper calls the root cause of the problem, and strong seeds no longer need heavy post-processing to be usable.","Since the framework requires only image-level labels and a pair of heterogeneous architectures, it inherits none of the scalability costs of external priors such as saliency maps, causal inference, or language supervision."],"supporting_citations":[{"why":"It supplies the empirical premise that ViTs and CNNs encode different internal visual representations, so the two architectures make complementary mistakes.","marker":"[43]"},{"why":"PCM is the strongest comparison baseline (33.56 mIoU), and the same pixel correlation module is inserted into the paper's ResNet teacher network.","marker":"[58]"},{"why":"It establishes the knowledge distillation paradigm on which the teacher-student transfer setup is built.","marker":"[22]"},{"why":"It inspires the learnable linear feature projection that maps ResNet and ViT features into a shared space before consistency is computed.","marker":"[55]"},{"why":"It justifies the choice of a linear rather than a larger projection layer for feature alignment.","marker":"[35]"},{"why":"It supplies the SAM-enhanced pseudo-label post-processing method that contributes to the final 52.93 mIoU.","marker":"[6]"},{"why":"It provides the frozen Segment Anything Model used at inference to refine CAM boundaries in that post-processing.","marker":"[27]"},{"why":"AffinityNet is a comparison baseline and post-processing method that fails on smoke, supporting the paper's point that class-agnostic affinity smoothing cannot fix low-contrast foregrounds.","marker":"[1]"},{"why":"Project RISE is one of the two out-of-distribution sources whose images are added to the training data to increase background diversity.","marker":"[24]"},{"why":"Smoke5K is the other source of supplementary out-of-distribution training images used to build the custom dataset.","marker":"[70]"}],"fun_headline_variants":["Cross-architecture consistency kills spurious correlations in weak segmentation","CNN + ViT team up to separate smoke from chimney without pixel labels","Teacher-student CNN-ViT beats co-occurrence bias, mIoU jumps to 47","No extra labels: Aligning CNN and ViT boosts WSSS from 33.6 to 47.4","Shared feature space suppresses chimney-smoke confusion in WSSS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the evaluation protocol: because only patches that contain visible smoke are scored, chimney-without-smoke patches — where residual spurious activation would show up — are never measured, so the reported mIoU gains do not by themselves prove that co-occurrence bias was suppressed.","fun_headline_variants_meta":{"raw":{"variants":["Cross-architecture consistency kills spurious correlations in weak segmentation","CNN + ViT team up to separate smoke from chimney without pixel labels","Teacher-student CNN-ViT beats co-occurrence bias, mIoU jumps to 47","No extra labels: Aligning CNN and ViT boosts WSSS from 33.6 to 47.4","Shared feature space suppresses chimney-smoke confusion in WSSS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00074,"raw_usage":{"total_tokens":3328,"prompt_tokens":992,"completion_tokens":2336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2229}},"tokens_in":608,"tokens_out":2336,"duration_ms":17097,"temperature":1.0,"reasoning_tokens":2229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:12:11.708493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Score the trained student model on chimney-only patches — chimney present, no visible smoke — and measure the smoke-class CAM activation: if a meaningful fraction of such patches light up, the spurious correlation is not suppressed and the reported mIoU, computed only on patches with visible smoke, would be hiding it. A companion check is to re-tune $\\lambda$, the CRF scaling, the SAM points-per-side, and the SAM fusion strategy on a held-out split of the same dataset; if the 47.37 and 52.93 figures shrink substantially, a portion of the gain reflects selection rather than the consistency mechanism.","supporting_citations":[{"cited_title":"Do vision transformers see like convolutional neural networks?","cited_arxiv_id":null,"evidence_quote":"It supplies the empirical premise that ViTs and CNNs encode different internal visual representations, so the two architectures make complementary mistakes."},{"cited_title":"Self-supervised equivariant attention mechanism for weakly super- vised semantic segmentation","cited_arxiv_id":null,"evidence_quote":"PCM is the strongest comparison baseline (33.56 mIoU), and the same pixel correlation module is inserted into the paper's ResNet teacher network."},{"cited_title":"A hybrid resnet-vit approach to bridge the global and local features for myocardial infarction detection","cited_arxiv_id":null,"evidence_quote":"It inspires the learnable linear feature projection that maps ResNet and ViT features into a shared space before consistency is computed."},{"cited_title":"Understanding the role of the projector in knowledge distillation","cited_arxiv_id":null,"evidence_quote":"It justifies the choice of a linear rather than a larger projection layer for feature alignment."},{"cited_title":"Segmentanything","cited_arxiv_id":null,"evidence_quote":"It provides the frozen Segment Anything Model used at inference to refine CAM boundaries in that post-processing."},{"cited_title":"Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation","cited_arxiv_id":"1803.10464","evidence_quote":"AffinityNet is a comparison baseline and post-processing method that fails on smoke, supporting the paper's point that class-agnostic affinity smoothing cannot fix low-contrast foregrounds."},{"cited_title":"Project RISE: recognizing industrial smoke emissions","cited_arxiv_id":null,"evidence_quote":"Project RISE is one of the two out-of-distribution sources whose images are added to the training data to increase background diversity."},{"cited_title":"Transmission-guided bayesian generative model for smoke segmentation","cited_arxiv_id":null,"evidence_quote":"Smoke5K is the other source of supplementary out-of-distribution training images used to build the custom dataset."}],"review_version":1}