{"id":"be4f60b6-74c5-4ea0-a41b-a0d3221c2b6a","arxiv_id":"2506.11142","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FARCLUSS is a semi-supervised segmentation method that blends fuzzy pseudo-labels, uncertainty weighting, class rebalancing, and contrastive prototypes, with modest benchmark gains.","lead":"FARCLUSS combines fuzzy top-K soft labels, entropy-based pixel weighting, per-batch class rebalancing, and a prototype contrastive loss into one semi-supervised segmentation pipeline. It reports small mIoU gains over prior methods on PASCAL VOC and Cityscapes, mainly at low labeled-data ratios.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own tables contradict the abstract's 'outperforms current state-of-the-art' claim: FARCLUSS trails CorrMatch, Diverse CoT, DDFP, or LogicDiag in multiple ResNet-101 settings.","rationale":"I agree with the reader that cross-paper comparability and missing error bars weaken the SOTA claim, and that the pseudocode/equations mismatch is a reproducibility issue. However, the strongest and most direct problem is internal: the paper's own tables show FARCLUSS losing to previously published methods in multiple settings. The reader's weakest assumption focuses on protocol differences between papers, but the contradiction exists even if every number in the tables is taken at face value. This makes the headline 'outperforms current state-of-the-art' not merely unverified but contradicted by the evidence presented. The component-level ablations are plausible and the method is a reasonable combination of known ideas, but the central claim of the abstract cannot stand as written. I therefore recommend rejecting the current version rather than accepting it conditionally, unless the authors substantially weaken the claims and reframe the contribution as a competitive, efficient alternative rather than a new SOTA.","tokens_in":16754,"tokens_out":6529,"duration_ms":69404,"concrete_test":"Perform a per-setting leaderboard tally from Tables 1-3: for each (dataset, backbone, label ratio) cell, compare FARCLUSS's mIoU strictly against every cited prior method in that cell. This is purely arithmetic on numbers already in the paper. If any prior method is strictly greater than FARCLUSS in a substantial fraction of ResNet-101 cells (e.g., more than 25%), the abstract's 'outperforms current state-of-the-art' claim is internally falsified and must be revised to 'competitive' before acceptance.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is empirical SOTA superiority. For this to hold, FARCLUSS should rank first in the reported comparisons. It does not, even within the paper's own tables. In Table 1 (ResNet-101, Pascal classic), FARCLUSS scores 76.4/78.2/79.0/80.3 at 1/16-1/2, while CorrMatch scores 76.4/78.5/79.4/80.6 and DDFP scores 75.0/78.0/79.5/81.2; at 1/4, Diverse CoT reaches 80.1 and at 1/2 DDFP reaches 81.2, both above FARCLUSS. In Table 2 (Pascal Blended, 513, ResNet-101), CorrMatch exceeds FARCLUSS at 1/8 (79.3 vs 79.0). In Table 3 (Cityscapes, ResNet-101), LogicDiag exceeds FARCLUSS at 1/8 (78.9 vs 78.5) and 1/4 (80.2 vs 80.0), and ties at 1/2 (81.0). Thus, across the high-backbone settings where the strongest baselines appear, FARCLUSS is frequently not the best method. The abstract's superiority claim is contradicted by the paper's own numbers, independent of any cross-paper protocol mismatch. This is load-bearing: if the claim is softened to 'competitive', the stated central contribution changes and the paper's novelty framing must be revised.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FARCLUSS is a semi-supervised semantic segmentation method built on a mean-teacher architecture. It proposes four components: (i) fuzzy soft pseudo-labels obtained by normalizing the teacher's top-K class probabilities; (ii) pixel-wise weights based on normalized entropy; (iii) batch-level adaptive class rebalancing using median-frequency inverse weights; and (iv) a prototype-based contrastive loss on high-confidence pixels. The paper reports mIoU comparisons on PASCAL VOC (classic and blended) and Cityscapes with ResNet-50/101 across 1/16, 1/8, 1/4, and 1/2 labeled splits, plus ablations and class-level analyses. The central claim is that FARCLUSS outperforms current state-of-the-art approaches, especially for under-represented classes and ambiguous regions.","tokens_in":17097,"tokens_out":11170,"duration_ms":108019,"significance":"If the empirical claims were fully supported, the framework would be a useful contribution: it targets two known weaknesses of semi-supervised segmentation (discarded uncertain pseudo-labels and class imbalance), and the individual loss terms in Eqs. (2)-(11) are coherent and computationally inexpensive. The ablation in Table 5 indicates that each component contributes, with fuzzy labeling having the largest effect. However, the significance is reduced by the fact that the headline state-of-the-art claim is not consistently supported by the paper's own tables: in several settings the method trails strong published baselines, most reported differences are below 1 mIoU point, and no variance estimates are provided. In addition, Algorithm 1 contains tensor-shape and batching errors that prevent reproduction. The paper's value is currently an incremental combination of existing ideas rather than a clearly demonstrated new state of the art.","major_comments":[{"comment":"The abstract's claim that the method 'outperforms current state-of-the-art approaches' is contradicted by the paper's own comparisons. In Table 1 (Pascal classic, ResNet-101), FARCLUSS scores 76.4/78.2/79.0/80.3 at 1/16, 1/8, 1/4, and 1/2, while CorrMatch scores 76.4/78.5/79.4/80.6, Diverse CoT reaches 80.1 at 1/4, and DDFP reaches 81.2 at 1/2. In Table 2 (Blended, train size 513, ResNet-101), CorrMatch is higher at 1/8 (79.3 vs. 79.0). In Table 3 (Cityscapes, ResNet-101), LogicDiag is higher at 1/8 (78.9 vs. 78.5) and 1/4 (80.2 vs. 80.0) and ties at 1/2 (81.0). Since the central contribution rests on empirical superiority, the authors should either soften the claim to 'competitive' or provide additional matched-protocol evidence that the method is actually ahead.","section":"Abstract; Tables 1-3"},{"comment":"Algorithm 1 is not runnable as written. The teacher output `pw` is produced from `xw` with batch size B, so `fuzzy` and `w_px` have batch dimension B, while `feat = g(torch.cat([xs1, xs2]))` has batch dimension 2B; consequently `criterion_ce(pred.log_softmax(1), fuzzy)` and the multiplication `w_px.view(-1,1,1,1) * w_cls[lbl].view(-1,1,1,1) * loss_u_map` have incompatible leading dimensions. The selection `sel = w_px > 0.5` cannot index `feat` for the same reason. The channel-masking step `pred = h(feat * mask)` is not described in Section 3, and `nn.CrossEntropyLoss()(f(x_l).log_softmax(1), y_l)` applies a log-softmax before a loss that already expects raw logits. Please correct these inconsistencies and state precisely how the weak view and the two strong views are batched.","section":"Algorithm 1"},{"comment":"The adaptive class rebalancing uses the median of the raw per-batch class counts `cnt`, including classes that do not occur in the batch. In small batches (e.g., Cityscapes with a 19-class label space), the median will frequently be zero, so `w_cls = 0/(cnt+1e-6)` becomes zero for every class that actually appears, driving the unsupervised loss to zero; when the median is positive, the weights remain highly sensitive to which classes were sampled. Please define the statistic over non-zero frequencies or use a smoothed count, and report the batch size and sensitivity to this choice.","section":"Eq. (7); Algorithm 1"},{"comment":"Most reported differences between FARCLUSS and the closest baselines are below 1 mIoU point (for example, Table 1, ResNet-50 1/16: 72.90 vs. 72.80 for CW-BASS; Table 3, ResNet-101 1/16: 77.2 vs. 77.3 for CorrMatch, where FARCLUSS is actually lower). No standard deviations, seeds, or reimplementations of baselines are reported, so the claimed advantage is not statistically distinguishable from noise. The authors should add variance estimates across multiple seeds and, where feasible, re-run the strongest baselines under the same training protocol.","section":"Tables 1-3"},{"comment":"The class-rebalancing implementation in Algorithm 1 does not match Eq. (8). Equation (8) multiplies each class term inside the KL divergence by `w_c`, whereas the algorithm computes `w_cls[argmax(fuzzy)]` and multiplies the entire per-pixel KL sum by this single weight. For a pixel whose fuzzy label distributes mass over two classes, which is the intended behavior of the top-K fuzzy labeling, the two formulations are quantitatively different. Please align the algorithm with Eq. (8) or explicitly define the argmax-based approximation and justify it.","section":"Eq. (8) vs. Algorithm 1"}],"minor_comments":[{"comment":"The implementation details state a normalized entropy threshold of 0.7, but the loss definitions in Eqs. (4)-(8) and Algorithm 1 do not refer to this threshold; please clarify where and how the threshold is applied.","section":"Section 4.2.4"},{"comment":"The caption says all methods are trained under the Full (1464) setting, while the text discusses semi-supervised settings; please specify the exact labeled fraction and training protocol for each qualitative comparison.","section":"Figure 3"},{"comment":"The class-level comparison introduces CPCL and DUEB without stating the training resolution, number of labeled images, or augmentation protocol for these runs; please add this information so the per-class numbers are interpretable.","section":"Table 6"},{"comment":"The text claims a 'monotonic descent' of the total loss and attributes spikes to bursts in L_u, but the figure lacks axis labels and the claims are not quantified; please add a properly labeled figure and describe whether the curves are smoothed.","section":"Section 4.5"},{"comment":"The text contains typos and notation glitches: 'that modulate' should be 'that modulates', several Greek symbols appear as 'Îż' in Algorithm 1, and venue names are inconsistent ('IJCNN' vs. 'IJCNN').","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main issue is claim calibration and reproducibility. The abstract and conclusion claim state-of-the-art superiority, but the paper's own tables show several settings where FARCLUSS trails published baselines, and Algorithm 1 contains batching and shape errors. These are fixable within the scope of a revision: the authors can soften the claims, correct the pseudocode, and add multi-seed variance. The novelty is largely a combination of known components (top-K soft labels, entropy weighting, median-frequency rebalancing, prototype contrast), so the journal should judge whether that combination is sufficient for publication once the empirical claims are made precise. There is no circularity concern: CW-BASS [5] is used as a baseline and is not the source of the claimed improvement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's own tables undercut its abstract. FARCLUSS trails CorrMatch, DDFP, Diverse CoT, and LogicDiag in several ResNet-101 settings, so the headline claim of outperforming current state-of-the-art is simply not supported by the numbers the authors report. That is the main thing to know.\n\nWhat is genuinely useful here is the integration. The combination of fuzzy top-K pseudo-labels (from Fuzzy Positive Learning), entropy-based pixel weighting, per-batch median class rebalancing, and a prototype contrastive term is not, as a package, in the existing literature. The per-batch median rebalancing is a small but real extension. The ablation study in Table 5 is informative: it shows each component contributes, with fuzzy labeling the largest single factor. The per-class results in Table 6 also give a reasonable picture of where the gains come from, and the unlabeled-data efficiency table is a nice addition.\n\nThe equations are coherent in isolation, but the soft spots are proportional to the claims. Most gains over strong baselines are under 1 mIoU. There are no error bars, no seeds, and no code. Algorithm 1 does not match the equations: the feature masking with a Bernoulli tensor is nowhere in the method text, and the way `sel` is applied to 4D features and 3D labels looks dimensionally wrong. More importantly, the paper's core fuzzy-labeling formula, Eq. (2)-(3), is borrowed directly from Fuzzy Positive Learning [47], yet that method is never compared in the experiments. That is a significant omission that weakens the novelty framing.\n\nWho is this for? Someone working in semi-supervised segmentation who wants a pragmatic combination of known tricks with a decent ablation study. It is not a breakthrough. It deserves a serious referee, but only if the authors are willing to compare against Fuzzy Positive Learning, report variance, fix the pseudocode, release code, and rewrite the abstract to say \"competitive or slightly superior\" rather than \"significant improvements.\" As it stands, the mismatch between claim and evidence is the biggest issue.","headline":"A plausible combination of known semi-supervised segmentation ideas whose own tables contradict the 'outperforms SOTA' claim; the ablations are solid but the paper needs major claim-softening and a direct comparison to its own fuzzy-labeling source.","tokens_in":17658,"tokens_out":2043,"would_cite":false,"duration_ms":24240,"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":"Semi-supervised segmentation can learn from uncertain pixels by keeping soft top-K labels and entropy-based weights, rather than discarding ambiguous predictions.","keywords":["semi-supervised learning","semantic segmentation","fuzzy pseudo-labeling","uncertainty estimation","class imbalance","contrastive learning","mean teacher","entropy weighting"],"falsifier":"Re-run FARCLUSS and its nearest baselines (especially UniMatch and PS-MT) under one shared codebase with identical augmentations, crop sizes, training lengths, and at least three random seeds, at the 1/16 Pascal VOC split; if the reported mIoU advantage of about 1 point does not reproduce or reverses, the central superiority claim fails.","tokens_in":16513,"feed_emoji":"🎯","tokens_out":4293,"duration_ms":39996,"temperature":0.7,"pith_summary":"This paper tries to establish that semi-supervised semantic segmentation performs better when a model uses, rather than discards, its uncertain predictions. It proposes FARCLUSS, a teacher–student framework whose unsupervised loss is built from fuzzy pseudo-labels: for each pixel it keeps the top-K class probabilities instead of a single hard label. Those soft labels are weighted pixel-wise by a reliability score derived from prediction entropy, reweighted per class by the inverse of that class's frequency in the current batch, and supplemented by a cheap prototype-based contrastive loss that pulls pixel embeddings toward class centroids. On Pascal VOC 2012 and Cityscapes, the authors report consistent mIoU improvements over comparison methods, with the largest advantages at low label ratios (1/16 and 1/8).","feed_headline":"Uncertain pixels become training data in new segmentation method","feed_subtitle":"FARCLUSS keeps soft top-K labels, entropy weighting, and batch class rebalancing to beat prior methods on Pascal VOC and Cityscapes.","key_machinery":"The load-bearing object is the fuzzy soft pseudo-label: a per-pixel distribution formed by taking the teacher's top-K class probabilities and renormalizing them, so each pixel keeps a small set of plausible classes instead of being forced into one hard label. The argument runs on three additional mechanisms attached to that label: an entropy-derived pixel weight $W_{h,w}=1-H(p^T_{:,h,w})$ that continuously down-weights ambiguous locations, a batch-level class weight $w_c=\\mathrm{median}(F)/(F_c+\\epsilon)$ that boosts rare classes, and a prototype contrastive loss that averages high-confidence embedded pixels per class and penalizes cosine distance from those centroids. Together they let the unsupervised loss carry soft, reliability-weighted, rebalanced supervision into the student model.","core_discovery":"The central discovery claimed is that uncertainty in teacher predictions is a reusable resource rather than noise to be filtered out. The paper's recipe is to transform each teacher probability map into a fuzzy label by renormalizing the probabilities of the top-K classes per pixel, then treat the remaining entropy as a continuous confidence signal. A per-pixel weight $W_{h,w}=1-H(p^T_{:,h,w})$ down-weights high-entropy pixels, a per-class weight $w_c=\\mathrm{median}(F)/(F_c+\\epsilon)$ amplifies rare classes in each batch, and a lightweight contrastive term on class prototypes regularizes feature geometry. The authors argue—and support with ablations—that the combination yields super-additive gains: removing fuzzy labeling costs about 6.2 mIoU on Pascal VOC 1/8, entropy weighting about 3.5, class rebalancing 2.6 on Cityscapes, and contrastive regularization 0.5–0.7.","pith_inferences":["The fuzzy pseudo-label can be read as a form of label smoothing that is adaptive per pixel, which might explain both the stability and the boundary-region improvements; one could test this by ablating top-K size K and comparing against fixed label smoothing.","The entropy weight $W_{h,w}=1-H$ acts as a soft curriculum that naturally up-weights easy pixels early and gradually admits harder ones, so a separate curriculum schedule may be unnecessary.","Per-batch inverse-frequency reweighting can be unstable in very small batches; a natural extension is to mix batch statistics with an exponentially averaged class-frequency estimate, which the paper does not explore.","The same four-component recipe could transfer to other dense prediction tasks, such as medical image segmentation, where boundary ambiguity and class imbalance are similarly dominant failure modes."],"forward_implications":["If the claimed gains hold, then strict confidence thresholding—the default in many FixMatch-style pipelines—is leaving useful signal on the table, and soft top-K labels can recover part of it.","At low label ratios (1/16 and 1/8), the method's reported edge over its closest baselines is largest, suggesting uncertainty-preserving supervision matters most when labeled data are scarce.","The per-batch class rebalancing scheme implies that class-imbalance compensation can be done locally and adaptively, without needing global class priors or manual tuning.","The prototype-based contrastive loss offers a low-overhead way to obtain feature compactness, pointing to a design where contrastive regularization is affordable in single-network teacher–student settings.","Because all components plug into a single shared teacher–student architecture, the method keeps the data requirements of streamlined methods like UniMatch while adding four corrective terms."],"supporting_citations":[{"why":"UniMatch is the primary weak-to-strong consistency baseline that FARCLUSS compares against and extends with fuzzy labeling and extra losses.","marker":"[3]"},{"why":"PS-MT is a strict-mean-teacher baseline with adaptive thresholds, serving as a key comparison for the claim that thresholding discards useful information.","marker":"[2]"},{"why":"U2PL supplies the unreliable-pseudo-label reinterpretation and the experimental setting for the blended Pascal VOC evaluation.","marker":"[29]"},{"why":"Mean Teacher is the underlying EMA teacher–student architecture that FARCLUSS adopts for pseudo-label generation.","marker":"[24]"},{"why":"Fuzzy positive learning provides the top-K renormalization idea that the paper's fuzzy pseudo-labeling directly builds on.","marker":"[47]"},{"why":"ReCo is the contrastive-learning baseline whose pairwise comparison cost FARCLUSS's prototype-based loss is designed to avoid.","marker":"[4]"},{"why":"CPS is the cross-pseudo-supervision method that FARCLUSS contrasts with, noting its uniform weighting of pseudo-labels.","marker":"[6]"}],"fun_headline_variants":["Uncertainty becomes a learning asset in segmentation","Fuzzy labels turn uncertain pixels into training data","Top-K fuzzy pseudo-labels beat strict thresholds","Entropy weighting and contrastive loss improve SSSS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains over prior methods are small—often below 1 mIoU point—and the comparison assumes identical training and evaluation settings across papers; if those settings actually differ, the ranking could flip.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty becomes a learning asset in segmentation","Fuzzy labels turn uncertain pixels into training data","Top-K fuzzy pseudo-labels beat strict thresholds","Entropy weighting and contrastive loss improve SSSS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3076,"prompt_tokens":918,"completion_tokens":2158,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":534,"tokens_out":2158,"duration_ms":17819,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:50:19.168378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run FARCLUSS and its nearest baselines (especially UniMatch and PS-MT) under one shared codebase with identical augmentations, crop sizes, training lengths, and at least three random seeds, at the 1/16 Pascal VOC split; if the reported mIoU advantage of about 1 point does not reproduce or reverses, the central superiority claim fails.","supporting_citations":[{"cited_title":"Revisiting Weak-to-Strong Consistency in Semi-Supervised Se- mantic Segmentation,","cited_arxiv_id":null,"evidence_quote":"UniMatch is the primary weak-to-strong consistency baseline that FARCLUSS compares against and extends with fuzzy labeling and extra losses."},{"cited_title":"Perturbed and strict mean teachers for semi-supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"PS-MT is a strict-mean-teacher baseline with adaptive thresholds, serving as a key comparison for the claim that thresholding discards useful information."},{"cited_title":"Semi-supervisedsemanticsegmentationusingunreliablepseudo- labels,","cited_arxiv_id":null,"evidence_quote":"U2PL supplies the unreliable-pseudo-label reinterpretation and the experimental setting for the blended Pascal VOC evaluation."},{"cited_title":"Mean teachers are better role models: Weight- averaged consistency targets improve semi-supervised deep learning results,","cited_arxiv_id":null,"evidence_quote":"Mean Teacher is the underlying EMA teacher–student architecture that FARCLUSS adopts for pseudo-label generation."},{"cited_title":"Fuzzy positive learning for semi-supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Fuzzy positive learning provides the top-K renormalization idea that the paper's fuzzy pseudo-labeling directly builds on."},{"cited_title":"Reco: Retrieve and co-segment for zero-shot transfer,","cited_arxiv_id":null,"evidence_quote":"ReCo is the contrastive-learning baseline whose pairwise comparison cost FARCLUSS's prototype-based loss is designed to avoid."},{"cited_title":"Semisupervised semantic segmentation with cross pseudo super- vision,","cited_arxiv_id":null,"evidence_quote":"CPS is the cross-pseudo-supervision method that FARCLUSS contrasts with, noting its uniform weighting of pseudo-labels."}],"review_version":1}