{"id":"23faf1c9-948f-44bf-b6a5-4012c2fa246c","arxiv_id":"2505.11075","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A teacher-student framework that separates class and mask confidence thresholds, uses CLIP to correct category pseudo-labels, and reweights mask loss by per-pixel uncertainty sets new state-of-the-art semi-supervised instance segmentation results.","lead":"This paper introduces a training recipe for semi-supervised instance segmentation that filters and corrects noisy computer-generated labels in three separate places: the instance, the category, and the pixel. It reports large gains on COCO and Cityscapes with little labeled data, including a +11.6 mAP improvement at 1% COCO labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DDTF rests on a mask-quality score (Eq. 4) that the paper's own ablation shows mis-filters small instances; the decoupling benefit is not established at the final training schedule.","rationale":"The reader's weakest assumption identifies the same soft spot: the teacher's confidence scores, specifically m_k, are treated as trustworthy mask-quality estimates. The paper itself concedes the score fails for small objects, and its ablation quantifies the harm (+0.3 AP_s when DDTF is removed). I considered whether CLIP's potential overlap with COCO images is a stronger threat, but DICC contributes only 0.8 mAP in the 10-epoch ablation and Appendix D shows CLIP has low recall on several common classes, so it is unlikely to be the main driver of the SOTA delta. The mask-quality issue is more load-bearing because DDTF is the paper's core decoupling contribution, and its benefit is not demonstrated at the final training length. The verdict remains CONDITIONAL: the empirical claim is plausible but provisional until the calibration of m_k is quantified and code or full-length ablations are provided.","tokens_in":17456,"tokens_out":17712,"duration_ms":186466,"concrete_test":"Run the COCO 1% setting for the full 368K-iteration schedule with DDTF disabled (i.e., m_t=0.0, keeping class threshold) and compare final mAP to the m_t=0.9 configuration. If the no-mask-filtering variant matches or exceeds mAP at convergence, the DDTF gain seen in the 10-epoch ablation is a short-horizon artifact. Additionally, on a small manually annotated holdout of unlabeled images, compute the Spearman correlation between m_k and mask IoU stratified by object area; if the correlation is near zero for small objects, the mask-quality score is not a valid IoU proxy for that regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DDFT's core assumption is that m_k, the mean over-0.5 sigmoid probability defined in Eq. 4, is a trustworthy proxy for mask IoU when filtering pseudo-labels. The paper's own Table 3 (COCO 1%, 10-epoch ablation) shows that removing DDTF and using coupled threshold 0.765 raises AP_s by +0.3 while lowering mAP by only 0.5. Thus m_t=0.9 actively discards good small-object masks, and the module's global gain is driven by medium/large objects. Sec. 4.3 and Appendix E explicitly concede that this fixed mask-quality threshold 'adversely affects small object quality' and warrants further investigation. Because all ablations are run for 10 epochs (73K iterations) while the headline SOTA numbers use 368K iterations, the module contribution at the reported operating point is unknown. If the mask-quality score is not calibrated by object size, the DDTF benefit may be an artifact of the short ablation schedule or the specific COCO size distribution, and the decoupling claim—the paper's central innovation—would not transfer to small-object-heavy regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PL-DC, a framework for semi-supervised instance segmentation built on a Mask2Former teacher-student setup with EMA. Three contributions are introduced: (i) Decoupled Dual-Threshold Filtering (DDTF), which filters instance pseudo-labels using separate thresholds on class confidence c_k and mask quality m_k instead of the coupled product; (ii) Dynamic Instance Category Correction (DICC), which uses CLIP to correct pseudo-label categories with a cosine schedule that anneals the CLIP weight from 0.5 to 0; and (iii) Pixel-level Mask Uncertainty-Aware loss (PMUA), which reweights the pixel-wise mask BCE by per-pixel uncertainty u_i = 1 - 2|sigma(q_i) - 0.5|. Experiments on COCO (1%, 2%, 5%, 10%, 100% labeled) and Cityscapes (5%, 10%, 20%, 30%) report consistent improvements over GuidedDistillation and other SSIS methods, including headline gains of +11.6 mAP at 1% COCO and +15.5 mAP at 5% Cityscapes relative to supervised Mask2Former. The paper also includes module ablations, hyperparameter sweeps, compatibility experiments with prior SSIS methods, and a short-schedule convergence analysis.","tokens_in":17677,"tokens_out":5599,"duration_ms":53429,"significance":"If the reported gains are reproducible, this is a meaningful contribution to SSIS: it identifies a real limitation of single-score pseudo-label filtering, introduces CLIP-based category correction to SSIS, and shows through the compatibility experiments in Table 5 that the proposed modules are not specific to one detector family. The framework is not circular: the category correction uses an external pretrained model, and the Appendix C computation is a gradient calculation rather than a derivation of the SOTA numbers. The main weakness is experimental validation: all ablations and threshold choices are made on a 10-epoch, 1%-COCO schedule, the headline numbers are single-run without error bars or released code, and the paper itself concedes that the mask-quality threshold mis-filters small objects. These gaps currently prevent the SOTA claim from being fully established.","major_comments":[{"comment":"The central DDTF claim is weakened by the paper's own small-object results. In Table 3, replacing DDTF with a coupled threshold of 0.765 raises AP_s by +0.3 mAP while lowering mAP by only 0.5, and the text in §4.3 and Appendix E explicitly states that the fixed mask-quality threshold 'adversely affects small object quality' and 'warrants further investigation.' Since Eq. (4) defines m_k as the mean sigmoid probability over confident foreground pixels and no size-dependent calibration is provided, the decoupling benefit is not established for small-object-heavy regimes. The authors should either introduce a size-adaptive mask-quality criterion or demonstrate that the AP_s regression disappears at the full training schedule.","section":"§4.3, Table 3; Appendix E"},{"comment":"All module ablations and hyperparameter sweeps (Tables 3 and 4) are run for 73K iterations (10 epochs) on 1% COCO, while the headline results use 360K iterations on COCO and 180K on Cityscapes. The convergence curves in Figure II are also limited to 10 epochs, and Appendix E states that the effect of removing DDTF on small objects reverses during the course of training. Therefore the contribution of each module at the final operating point is unknown, and the attribution of the final SOTA gains to the three modules is not directly supported. The ablations should be repeated at the full schedule or at a substantially longer schedule to validate the modules' contributions.","section":"§4.1 and §4.3"},{"comment":"All reported numbers are single-run, with no error bars, seeds, or statistical significance tests. Several ablated differences in Table 4 are on the order of 0.4-0.8 mAP (e.g., c_t = 0.8 vs. 0.85, and alpha = 0.999 vs. 0.9996), which is within a plausible range of run-to-run variation for a short training with batch size 4; the same concern applies to the module differences in Table 3. Reporting variance over at least three seeds, or releasing code and logs so that the headline numbers can be reproduced, is necessary to support the claim of 'new state-of-the-art results.'","section":"§4.2, Tables 1-2"},{"comment":"The thresholds m_t = 0.9 and c_t = 0.85 are selected by tuning on the same 1%-COCO, 10-epoch protocol that is used for the headline ablations, and are then applied to all COCO ratios, all Cityscapes ratios, and the full 360K/180K schedules without held-out validation. Table 4(a) shows that the selected thresholds are not robust in their neighborhood: both lower combinations (0.8/0.7) and the higher combination (0.9/0.9) produce notably worse mAP. The paper should either validate the thresholds on a separate split or schedule, or provide sensitivity analyses at the final training schedule, to rule out threshold-selection effects on the reported gains.","section":"§4.3, Table 4"}],"minor_comments":[{"comment":"The heading 'Abalation Study' should be 'Ablation Study', and 'Superised' in Table 1 is a typo for 'Supervised.'","section":"§4.3 and Table 1"},{"comment":"When no pixel satisfies sigma(q_i^k) > 0.5, the denominator in the mask-quality score m_k is zero; the authors should specify a default value for degenerate mask predictions.","section":"§3.1, Eq. (4)"},{"comment":"The claim that the coupled instance score is not positively correlated with IoU is supported only by a qualitative scatter plot; please include Pearson or Spearman correlation coefficients for panels (a), (b), and (c).","section":"Figure 2"},{"comment":"The appendix is labeled a 'validity proof' but contains only a chain-rule computation of the gradient; since the derivation assumes the linear model t = theta x, which is not the actual Mask2Former output, the authors should present this as a simplified local analysis rather than a proof of robustness.","section":"Appendix C"},{"comment":"Each row of Table 5 (a-h) is described in a single sentence; adding a column that states the exact modification, and the component it replaces, would make the compatibility claims less ambiguous.","section":"§4.4, Table 5"},{"comment":"The text says the CLIP weight w decays from 0.5 to 0, and Eq. (7) uses a cosine schedule; please confirm the schedule's exact range and state whether 'it_max' refers to the full 360K iterations or to the current training stage.","section":"§3.2, Eq. (7)"},{"comment":"The CLIP precision/recall analysis uses ground-truth masks from COCO val; please state whether the same text template is used for all 80 categories and whether prompt ensembling is applied.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core ideas are reasonable, but the experimental validation currently lags behind the strength of the claims. The most important issue is the mismatch between the 10-epoch ablation/tuning protocol and the 50-epoch headline numbers; I would ask the editor to require full-schedule ablations or at minimum multi-seed variance reporting before acceptance. The authors also concede a small-object weakness in their central DDTF module, which should be addressed rather than deferred."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's what you need to know: PL-DC is a serious systems paper that improves semi-supervised instance segmentation at low label ratios, and the dynamic CLIP-based category correction (DICC) is the genuinely new idea. The state-of-the-art claims are plausible but not established—there's no code, no error bars, and the key filtering module has a known small-object weakness that the authors themselves flag.\n\nWhat's good: The paper decomposes pseudo-label noise into instance, category, and pixel levels, and addresses each with a clean module. DDTF decouples mask/class thresholds, DICC fuses teacher and CLIP probabilities with a cosine-decayed weight, and PMUA reweights mask loss by pixel uncertainty. The ablations cover all three, and the compatibility experiments show DDTF/PMUA/DICC also improve GuidedDistillation and PAIS. That's solid engineering with honest reporting, including a candid appendix about CLIP's per-category recall.\n\nSoft spots: The stress-test concern is fair. Eq. (4)'s mask-quality score is a mean foreground sigmoid, which does not track IoU well for small objects. The paper's own Table 3 shows that removing DDTF actually increases AP_s by 0.3 while lowering mAP by 0.5, and the text concedes the fixed threshold 'adversely affects small object quality' and 'warrants further investigation.' Because all ablations run for 10 epochs (73K iterations) while the headline tables use 360K, we don't know whether DDTF's contribution is an artifact of the short schedule. That's a genuine gap, not a manufactured one. Additionally, the results are single-run, code isn't available yet, and key hyperparameters (m_t, c_t, lambda, alpha) are tuned on the same protocol as the headline numbers. Those issues are all fixable in revision.\n\nWho it's for: SSIS researchers will want to read this, and the DICC idea is worth borrowing for other dense prediction tasks. It deserves a thorough peer review. My recommendation: major revision asking for code release, multi-seed results, and a size-aware analysis of DDTF at the full training schedule. If those come through, the deltas will likely hold.","headline":"A credible SSIS framework whose CLIP-based category correction is the real novelty, but the SOTA deltas are provisional given no code, no seeds, and a known small-object bias in DDTF.","tokens_in":18239,"tokens_out":4428,"would_cite":true,"duration_ms":42547,"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":"This paper claims that decoupling class-quality and mask-quality thresholds for pseudo-label filtering, adding CLIP-based category correction, and reweighting mask loss by pixel uncertainty together set new state-of-the-art results in…","keywords":["semi-supervised instance segmentation","pseudo-label filtering","dual-threshold decoupling","class quality","mask quality","CLIP category correction","pixel uncertainty weighting","teacher-student learning"],"falsifier":"Measure the Spearman rank correlation between the teacher's mask-quality score $m_k$ (Eq. 4) and the IoU of the predicted mask against the ground-truth mask, computed on a held-out set of unlabeled COCO images with oracle masks, and repeat for small, medium, and large objects; if for small objects the correlation is near zero or negative, DDTF's fixed $m_t=0.9$ would filter out precisely the best small-instance pseudo-labels, and the claimed small-object gains would fail to reproduce.","tokens_in":17222,"feed_emoji":"🧩","tokens_out":10075,"duration_ms":84930,"temperature":0.7,"pith_summary":"The paper targets semi-supervised instance segmentation, where a model is trained on a few labeled images together with many unlabeled ones to produce object masks and their categories. Its central claim is that pseudo-label filtering should not use a single threshold on the product of class confidence and mask confidence, because that product forces a trade-off between classification quality and mask quality; instead the two qualities should be gated independently. The proposed framework couples this decoupled dual-threshold filter with a CLIP-based category-correction module whose influence fades as the teacher improves, and with a per-pixel uncertainty weighting that down-weights ambiguous mask predictions in the loss. The authors report state-of-the-art results on COCO and Cityscapes at every tested label ratio, with the largest gains under extreme scarcity (+11.6 mAP at 1% COCO labels and +15.5 mAP at 5% Cityscapes labels). If these results hold, the practical bottleneck in semi-supervised instance segmentation is not the amount of pseudo-label data but its quality control.","feed_headline":"Two thresholds beat one: +15.5 mAP in semi-supervised segmentation","feed_subtitle":"Separating class and mask confidence gates, plus CLIP corrections and uncertainty weighting, explains the gains.","key_machinery":"The central object is the decoupled dual-threshold filter: rather than one threshold on the coupled score $s_k=c_k \\cdot m_k$, the teacher's pseudo-labels are accepted only when class confidence $c_k \\ge c_t$ and mask confidence $m_k \\ge m_t$ independently, with $m_t=0.9$ and $c_t=0.85$. The mask confidence $m_k$ is the mean sigmoid foreground probability over pixels predicted as foreground, defined in Eq. 4. Two mechanisms complete the pipeline: Dynamic Instance Category Correction (DICC) blends the teacher's class distribution with CLIP's zero-shot patch-level class distribution using a weight that decays from 0.5 to 0, and Pixel-level Mask Uncertainty-Aware (PMUA) weights each pixel's mask loss by $1-u_i^k$, where $u_i^k = 1 - 2|\\sigma(q_i^k)-0.5|$ captures how near the teacher's prediction is to the decision boundary. DDTF is the entry gate that determines which instances enter the training pool, DICC repairs their category labels, and PMUA suppresses the noisy pixel-level supervision inside the kept masks.","core_discovery":"At the heart of the paper is the observation that an instance pseudo-label carries two independent qualities - a class quality (how confident the model is about the category) and a mask quality (how confident it is about the pixel-level foreground) - and that their product, the standard instance score, does not reliably rank pseudo-labels by true mask IoU. The authors argue that filtering with a single threshold on this product sets up a competition between the two qualities, so they propose the Pseudo-Label Quality Decoupling and Correction (PL-DC) framework. Its Decoupled Dual-Threshold Filtering keeps a pseudo-label only when the class confidence meets $c_t=0.85$ and the mask confidence meets $m_t=0.9$ independently. Its Dynamic Instance Category Correction module feeds the masked image patch to CLIP, fuses CLIP's zero-shot category distribution with the teacher's predicted distribution under a cosine-decayed weight that starts at 0.5 and falls to 0, and takes the argmax as the corrected category. Its Pixel-level Mask Uncertainty-Aware loss reweights each pixel's mask cross-entropy by $1-u_i^k$, where $u_i^k$ is highest when teacher confidence is near 0.5. The paper claims that this three-part combination produces new state-of-the-art results in semi-supervised instance segmentation on COCO and Cityscapes, exceeding the previous best method at all evaluated labeled-data ratios.","pith_inferences":["A natural extension the authors do not test: applying the same decoupling logic to the mask-quality score by learning a per-size or per-instance mask threshold rather than a fixed $m_t=0.9$, since their own ablation shows small objects are hurt by the fixed mask gate.","The same three-step treatment (separate gates, external-class correction, per-pixel confidence weighting) could transfer to other dense prediction tasks built on teacher-student pipelines, such as semi-supervised panoptic segmentation or video instance segmentation.","The DICC module's cosine decay assumes the teacher's class distribution monotonically improves; a testable consequence is that the optimal decay schedule should depend on per-category teacher accuracy, and a category-aware schedule might extend the gains.","The paper's evidence that the coupled score does not track IoU suggests a broader diagnostic: in any semi-supervised dense-prediction setup, one should verify that the confidence score used to filter pseudo-labels actually correlates with the ground-truth quality metric, not just with the teacher's self-assessment."],"forward_implications":["If the results are right, the standard practice of a single confidence-product threshold in semi-supervised instance segmentation is suboptimal, and independent class and mask gates should be the default.","Extreme label scarcity becomes less forbidding: the reported +11.6 mAP at 1% COCO and +15.5 mAP at 5% Cityscapes imply that well-filtered pseudo-labels can nearly close the gap to much larger labeled sets.","The compatibility experiments indicate the category-correction and uncertainty modules are drop-in additions that improve other teacher-student SSIS frameworks as well, not just the paper's own combination.","Because all three modules act only during training, the inference-time cost of the model is unchanged; any gains come without extra test-time computation."],"supporting_citations":[{"why":"The previous state-of-the-art method that filters pseudo-labels with a single coupled score; it is the main baseline on COCO and Cityscapes.","marker":"[3]"},{"why":"Mask2Former, the query-based instance segmentation network that serves as the teacher and student backbone; the paper builds PL-DC on top of it.","marker":"[11]"},{"why":"A prior method that already used dual-threshold filtering but via a learned mask-IoU prediction branch; the paper compares DDTF against it.","marker":"[23]"},{"why":"A prior SSIS method that reweights mask loss with a boundary-preserving map; the paper compares PMUA against it.","marker":"[41]"},{"why":"CLIP, the vision-language model whose zero-shot category predictions power the Dynamic Instance Category Correction module.","marker":"[37]"},{"why":"COCO, the primary benchmark dataset used for most experiments and ablations.","marker":"[30]"},{"why":"Cityscapes, the autonomous-driving dataset used to test generalization of the framework.","marker":"[13]"}],"fun_headline_variants":["Decoupled thresholds and CLIP correction yield +15.5 mAP SSIS","Two thresholds, not one: +11.6 mAP with 1% COCO","Separating class and mask scores boosts semi-supervised segmentation","Dual-threshold decoupling with CLIP fixes pseudo-label quality","Decoupling quality in pseudo-labels: +15.5 mAP on Cityscapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a teacher model's own confidence scores - especially the mask-quality score $m_k$, the mean foreground sigmoid probability - are honest estimates of pseudo-label quality on unlabeled data; if that confidence-to-quality link breaks, as the paper concedes it already does for small objects, the decoupled thresholds either admit noisy masks or discard good ones, and the reported gains are not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Decoupled thresholds and CLIP correction yield +15.5 mAP SSIS","Two thresholds, not one: +11.6 mAP with 1% COCO","Separating class and mask scores boosts semi-supervised segmentation","Dual-threshold decoupling with CLIP fixes pseudo-label quality","Decoupling quality in pseudo-labels: +15.5 mAP on Cityscapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3698,"prompt_tokens":1133,"completion_tokens":2565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":749,"tokens_out":2565,"duration_ms":17688,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:58:05.765844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the Spearman rank correlation between the teacher's mask-quality score $m_k$ (Eq. 4) and the IoU of the predicted mask against the ground-truth mask, computed on a held-out set of unlabeled COCO images with oracle masks, and repeat for small, medium, and large objects; if for small objects the correlation is near zero or negative, DDTF's fixed $m_t=0.9$ would filter out precisely the best small-instance pseudo-labels, and the claimed small-object gains would fail to reproduce.","supporting_citations":[{"cited_title":"Guided distillation for semi-supervised instance segmentation","cited_arxiv_id":null,"evidence_quote":"The previous state-of-the-art method that filters pseudo-labels with a single coupled score; it is the main baseline on COCO and Cityscapes."},{"cited_title":"Schwing, Alexan- der Kirillov, and Rohit Girdhar","cited_arxiv_id":null,"evidence_quote":"Mask2Former, the query-based instance segmentation network that serves as the teacher and student backbone; the paper builds PL-DC on top of it."},{"cited_title":"Pseudo-label alignment for semi-supervised instance segmentation","cited_arxiv_id":null,"evidence_quote":"A prior method that already used dual-threshold filtering but via a learned mask-IoU prediction branch; the paper compares DDTF against it."},{"cited_title":"Noisy bound- aries: Lemon or lemonade for semi-supervised instance seg- mentation?2022 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 16805–16814,","cited_arxiv_id":null,"evidence_quote":"A prior SSIS method that reweights mask loss with a boundary-preserving map; the paper compares PMUA against it."},{"cited_title":"Learning transferable visual models from natural language supervi- sion","cited_arxiv_id":null,"evidence_quote":"CLIP, the vision-language model whose zero-shot category predictions power the Dynamic Instance Category Correction module."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"COCO, the primary benchmark dataset used for most experiments and ablations."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding","cited_arxiv_id":null,"evidence_quote":"Cityscapes, the autonomous-driving dataset used to test generalization of the framework."}],"review_version":1}