{"id":"c029cb3c-f5a9-490e-ab8e-892930fb5544","arxiv_id":"2504.12599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"3DResT applies teacher-student semi-supervised learning to 3D referring expression segmentation, promoting high-agreement pseudo-labels into the labeled set and dynamically weighting low-agreement ones, and reports a +8.34 mIoU gain over fully supervised training at 1% labels on ScanRefer.","lead":"This paper introduces 3DResT, the first semi-supervised method for 3D referring expression segmentation, which uses a teacher-student setup to learn from many unlabeled point cloud scans. It reports large gains over fully supervised training when only 1% of labels are available, potentially cutting annotation costs for 3D scene understanding.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Teacher–student IoU is a self-referential quality signal: TSCS/QDW may select and upweight confidently wrong masks, so the core mechanism is unvalidated.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: teacher–student agreement is used as a pseudo-label quality signal despite the teacher being an EMA of the student. This is the most serious issue because both novel modules (TSCS and QDW) depend on it. If the signal is uninformative, the paper's stated mechanism fails, and the observed improvements may have a different, simpler explanation. The concern is not fatal at this stage because the empirical results are coherent and the issue is directly testable. I therefore maintain the CONDITIONAL verdict: the paper should provide evidence that teacher–student IoU tracks true mask quality, ideally by comparing with ground truth on a subset, before the claimed mechanism is accepted. The absence of error bars and unreleased code add to the need for such validation but do not change the verdict category.","tokens_in":13742,"tokens_out":6801,"duration_ms":71113,"concrete_test":"On a randomly held-out subset of unlabeled ScanRefer scenes with ground-truth masks, compute for each sample (i) the teacher–student mask IoU used by TSCS and QDW, and (ii) the IoU between the teacher pseudo-label and the ground-truth mask. If the Spearman correlation between these two quantities is low (e.g., below 0.3) or if the average true IoU of pseudo-labels passing the 0.95 teacher–student threshold is below 0.8, then the quality signal is invalid and the core mechanism of TSCS/QDW is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TSCS and QDW improve semi-supervised 3D-RES rests on the assumption that the IoU between the student prediction and the teacher pseudo-label (Eq. 6 for TSCS, Eq. 7 for QDW) is a trustworthy estimate of pseudo-label correctness. Because the teacher is an EMA of the student (Eq. 5), the two predictors are strongly correlated; high agreement can simply mean the model is confidently wrong. The paper provides no direct evidence that this agreement correlates with ground-truth mask quality. Thus, TSCS's threshold of 0.95 could admit many incorrect masks into the labeled set, and QDW could upweight those same errors. The empirical ablations show only that adding these modules helps, not that the selection criterion is actually selecting high-quality pseudo-labels. The reported gains could partly stem from increased labeled-set volume or implicit regularization rather than from the proposed quality mechanism. This is a load-bearing flaw because the paper's explanation of why the modules work hinges on the validity of this self-referential quality signal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 3DResT, which it presents as the first semi-supervised learning framework for 3D referring expression segmentation (3D-RES) on point clouds. The method adopts a teacher–student architecture where the teacher is an EMA of the student, generates pseudo-labels for unlabeled data, and supervises the student through a combined supervised and unsupervised loss. Two new components are proposed: Teacher-Student Consistency-Based Sampling (TSCS), which measures the IoU between the teacher pseudo-label and the student prediction and promotes pseudo-labels with IoU above a threshold (0.95) into the labeled set; and Quality-Driven Dynamic Weighting (QDW), which uses the same teacher–student IoU as a per-sample weight on the unsupervised loss so that low-quality pseudo-labels are down-weighted rather than discarded. Experiments on ScanRefer at 1%, 2%, 5%, and 10% labeled data show consistent mIoU improvements over a fully supervised baseline and over RefTeacher, a 2D referring-expression SSL method adapted to 3D. Ablations in Table II indicate that both TSCS and QDW contribute, with TSCS providing the larger gain.","tokens_in":14021,"tokens_out":4349,"duration_ms":46212,"significance":"If the results hold, this is a valuable contribution: it opens a new task setting (semi-supervised 3D-RES) and reports large gains in low-label regimes (e.g., +8.34 mIoU at 1% labels, Table I), which is practically relevant given the high cost of instance-level 3D annotations. The method itself is simple and builds on established teacher–student ideas, making it a reasonable baseline for future work. The paper also includes ablations that show both proposed modules help, and Table IV provides some evidence of the sensitivity of TSCS to the sampling period. However, the central explanatory claim—that teacher–student agreement is a trustworthy proxy for pseudo-label correctness—is not directly validated, and several implementation details (notably the sampling period K) are unspecified. The empirical gains are consistently positive across label ratios, which partially mitigates the concern, but the lack of uncertainty estimates and the small margins in some comparisons (e.g., Table III) leave the strength of the claim somewhat fragile.","major_comments":[{"comment":"The core mechanism is that teacher–student IoU, computed in Eq. (6) for TSCS and Eq. (7) for QDW, measures pseudo-label quality. Because the teacher is the EMA of the student (Eq. 5), the two models are strongly correlated, so high agreement can occur even when both are confidently wrong. The paper provides no direct evidence that this agreement correlates with ground-truth mask quality: there is no analysis of the precision of selected pseudo-labels, no comparison of TSCS selection against random selection or confidence-based selection of the same set size, and no study of whether QDW's per-sample weights track actual correctness. The ablation in Table II shows that the modules improve performance, but it does not show that the quality-selection mechanism is the reason. I ask the authors to add an experiment that measures the agreement between the teacher–student IoU and the IoU against ground truth at selection time, or to compare TSCS against a control that selects the same number of pseudo-labels randomly. Without such validation, the attribution of the gains to 'quality' is not supported.","section":"Algorithm 1, Section IV.A"},{"comment":"The paper does not specify the sampling period K used in the main experiments (Section IV.A lists alpha, s, and lambda_u but not K). Since Table IV shows that the choice of sampling period materially affects performance (1.61 mIoU difference between selecting at Mid vs. not selecting), omitting K prevents reproduction of the main results. Please state the default K and clarify whether sampling is a one-time event or periodic.","section":"Algorithm 1, Section IV.A"},{"comment":"The main results and ablations are presented without error bars or significance tests, even though some differences are small (e.g., Table III QDW vs. baseline, +0.15 mIoU overall). Reporting the mean and standard deviation over multiple seeds (e.g., 3 runs) for the key comparisons would make the claims more robust. Since the paper's headline claim is an '8.34-point improvement', establishing that this is not a single-run artifact is important.","section":"Tables I–IV, Section IV"},{"comment":"Table III's setup is ambiguous: the 'Baseline' row is not defined (does it include TSCS? does it include QDW?), and the 'predicted-weight' row is said to be RefTeacher in the text but the table does not say so. Since the comparison is meant to justify QDW over an alternative, the exact configuration of each row must be explicit.","section":"Table III, Section IV-D"}],"minor_comments":[{"comment":"The term 'correlation score' is used for what is simply the IoU between two binary masks; it is not a statistical correlation. Consider renaming it 'consistency score' or 'IoU score' for clarity.","section":"Eq. (6)"},{"comment":"Eq. (7) defines a per-sample weight lambda_{unsup}^{weight}, but Eq. (11) writes the unsupervised loss as a single equation with lambda_{unsup}^{weight} multiplying the full BCE and Dice terms. Please clarify whether the weighting is applied per-sample and then averaged over the batch, and give the exact averaging formula.","section":"Eqs. (7) and (11)"},{"comment":"The caption of Table IV contains a typo: 'diffenet' should be 'different'. Also, the row labels 'Early', 'Mid', 'Late' should be defined in terms of epochs or iterations so that the reader can relate them to the unspecified K.","section":"Table IV"},{"comment":"Algorithm 1 uses 'epoch = K' as the sampling condition, but the training loop is described in iterations. Please clarify the relationship between epochs and iterations, and specify how many epochs correspond to the burn-in stage (N1=2k steps) and the total training length.","section":"Section IV-A and Algorithm 1"},{"comment":"The Index Terms line has a formatting issue: 'T erms' should be 'Terms'.","section":"Index Terms"},{"comment":"The caption says 'The correl in TSCS is a metric...' but the figure itself uses 'correl' without a clear definition. After introducing Eq. (6), the figure could be updated to label the metric as 'IoU' for consistency.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core empirical finding—consistent gains in semi-supervised 3D-RES—is likely of interest. The main risk is that the proposed mechanisms are justified by a quality proxy that is never validated against ground truth. I do not doubt the experiments as reported, but the lack of error bars and the unspecified sampling period K are reproducibility issues that should be fixed before acceptance. The claim of being 'the first' semi-supervised 3D-RES framework should also be checked against more recent literature, as the paper only compares against a 2D method (RefTeacher) and does not adapt other 3D SSL methods from segmentation or detection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it is the first to apply semi-supervised learning to 3D referring expression segmentation, and it makes a solid empirical case that a teacher-student setup helps a lot in low-label regimes. On 1% of ScanRefer labels, the method improves mIoU by 8.34 points over the fully supervised baseline. The two proposed pieces, TSCS and QDW, are simple and plausible: TSCS promotes high-agreement teacher-student pseudo-labels into the labeled set, and QDW down-weights low-agreement ones instead of discarding them. Ablations show both contribute, with TSCS giving the bigger boost. That is a reasonable contribution, not a revolution.\n\nThe soft spots are real but not fatal. The stress-test concern about the self-referential quality signal is fair: because the teacher is an EMA of the student, IoU between their predictions is not an independent oracle, and high agreement could mean confidently wrong. The paper would be much stronger with a direct precision analysis, e.g., what fraction of masks that pass the 0.95 threshold actually match ground truth. That is a validation gap, though, not a demonstrated flaw. The empirical gains suggest the mechanism is doing something useful, even if the explanation is not fully pinned down.\n\nOther issues are more mundane. All results are single-run with no error bars or significance tests. The sampling period K is an input in Algorithm 1 but no default value is given in the experimental settings; Table IV uses vague stages instead of concrete numbers. No code is released, which limits reproducibility. The evaluation is on ScanRefer only, so we do not know if the gains generalize to Nr3D or Sr3D. These are addressable in revision.\n\nI would send this to peer review. The core finding is useful, the experiments are coherent, and the weaknesses are mostly about validation and reporting rather than a load-bearing contradiction. If the authors add a pseudo-label quality analysis, release code, and standard deviations, the paper becomes a solid baseline for future work in 3D-RES and related SSL tasks. I would cite it if I worked in this area.","headline":"First semi-supervised 3D-RES baseline with two sensible but under-validated components; deserves peer review with ask for stronger evidence on pseudo-label quality.","tokens_in":14534,"tokens_out":1709,"would_cite":true,"duration_ms":19651,"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 3DResT claims that 1% of labels can beat full supervision in 3D referring segmentation by recycling pseudo-labels.","keywords":["3D referring expression segmentation","semi-supervised learning","teacher-student framework","pseudo-label","point cloud","consistency-based sampling","dynamic weighting","ScanRefer"],"falsifier":"Hold out the true masks for the unlabeled split and measure the correlation between Eq. (6)'s teacher-student IoU and the true IoU of the teacher pseudo-label; if that correlation is near zero or negative, the sampling and weighting cannot be the mechanism behind the reported 8.34-point gain.","tokens_in":13555,"feed_emoji":"🎯","tokens_out":8287,"duration_ms":73635,"temperature":0.7,"pith_summary":"3D referring expression segmentation cuts out the object in a 3D point cloud that a sentence describes, and labeling such masks is expensive. This paper introduces 3DResT, the first semi-supervised teacher-student framework for this task, which keeps the cheap language descriptions and drops the masks for unlabeled scenes. Its central claim is that on the ScanRefer benchmark, training with only 1% of the labeled data reaches 25.41 mean IoU, beating the fully supervised baseline at 17.07 by 8.34 points. The paper argues that existing semi-supervised pipelines waste pseudo-labels: good ones are used only as weak auxiliary losses, and bad ones are discarded outright. 3DResT instead promotes teacher-student-agreed masks into the labeled set and keeps the rest with weights proportional to their agreement.","feed_headline":"1% of labels beats full supervision in 3D referring segmentation","feed_subtitle":"Teacher-student agreement promotes good pseudo-labels and weights the rest, lifting mean IoU from 17.07 to 25.41.","key_machinery":"The load-bearing object is the teacher-student agreement score $correl$, defined as the IoU between the student's mask prediction $Y_s^u$ and the teacher's pseudo-label $\\hat{Y}_t^u$ (Eq. 6). It is used twice: TSCS selects pseudo-labels with $correl > 0.95$ and merges them into the labeled dataset, while QDW sets each unlabeled sample's loss weight to $\\lambda_{\\text{weight}}^{\\text{unsup}} = \\mathrm{IoU}(Y_s^u, \\hat{Y}_t^u)$ (Eq. 7), so low-quality pseudo-labels keep contributing with reduced influence. Because the teacher's weights are the exponential moving average of the student's (Eq. 5), the two networks provide comparable predictions, and the agreement score serves as a stand-in for pseudo-label correctness.","core_discovery":"The paper's claim is that pseudo-labels themselves, not the amount of labeled data, are the bottleneck in semi-supervised 3D-RES. It proposes to treat a pseudo-label as trustworthy ground truth when the student's prediction and the teacher's pseudo-label agree above a high IoU threshold of 0.95, adding that sample to the labeled set via Teacher-Student Consistency-Based Sampling (TSCS). All other pseudo-labels are retained rather than filtered, with each unlabeled sample's loss weighted by the teacher-student IoU through Quality-Driven Dynamic Weighting (QDW). On ScanRefer, this combination yields 25.41 mIoU with 1% labeled data, an 8.34-point improvement over fully supervised training, and the gain grows as the labeled fraction shrinks.","pith_inferences":["The same promote-and-softly-weight recipe could transfer to other language grounding tasks where each annotation supervises a single object, such as 2D referring segmentation or 3D referring comprehension.","Because the labeled set only grows, early promoted pseudo-labels could become stale as the teacher improves; a testable extension is to periodically re-score and evict previously promoted samples.","Substituting the teacher-student IoU weight in QDW with a calibrated confidence estimate would test whether agreement is really the right proxy, or whether the gain comes from sample reweighting in general."],"forward_implications":["With 1% of ScanRefer labels, 3DResT reaches 25.41 mIoU versus 17.07 for full supervision, and the margin grows as the labeled fraction shrinks from 10% down to 1%.","Selecting pseudo-labels at the middle of teacher-student training works better than selecting them early or late, so the sampling period is a tunable lever for performance.","Ablations attribute most of the gain to TSCS: selecting high-agreement pseudo-labels adds 1.61 mIoU points at 1% labels, while QDW alone adds 0.15.","Keeping low-quality pseudo-labels with soft weights outperforms discarding them or weighting them by the model's own predicted confidence scores.","Unlabeled data for the framework need only language descriptions, not masks, so new unlabeled material can be collected from existing vision-language sources without annotation."],"supporting_citations":[{"why":"Supplies the ScanRefer benchmark, evaluation metric, and the labeled/unlabeled splits used for the 1%, 2%, 5%, and 10% regimes.","marker":"[21]"},{"why":"Provides the base 3D-RES model shared by all compared methods, including the supervised loss terms.","marker":"[14]"},{"why":"The 2D referring semi-supervised baseline that contributes the burn-in stage, EMA teacher update, and attention-alignment loss used in 3DResT.","marker":"[46]"},{"why":"Underpins the EMA teacher update that prevents the teacher from overfitting to limited labeled data.","marker":"[51]"},{"why":"Provides the ScanNet scenes from which the ScanRefer dataset is built.","marker":"[16]"},{"why":"Supplies the Dice loss used in both the supervised and unsupervised segmentation objectives.","marker":"[48]"}],"fun_headline_variants":["Pseudo-labels, not data, are the bottleneck in 3D-RES","Semi-supervised 3D-RES: 1% labels outdo full supervision","Teacher-student agreement boosts 3D segmentation with 1% labels","3DResT: Rethink pseudo-labels for semi-supervised 3D-RES","Quality-weighted pseudo-labels lift 3D-RES mIoU by 8.34 pts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that when the student and teacher produce nearly identical masks, the teacher's pseudo-label is accurate enough to be treated as ground truth, even though the teacher is just an averaged copy of the student and both could be confidently wrong together.","fun_headline_variants_meta":{"raw":{"variants":["Pseudo-labels, not data, are the bottleneck in 3D-RES","Semi-supervised 3D-RES: 1% labels outdo full supervision","Teacher-student agreement boosts 3D segmentation with 1% labels","3DResT: Rethink pseudo-labels for semi-supervised 3D-RES","Quality-weighted pseudo-labels lift 3D-RES mIoU by 8.34 pts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001084,"raw_usage":{"total_tokens":4585,"prompt_tokens":1052,"completion_tokens":3533,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":3418}},"tokens_in":668,"tokens_out":3533,"duration_ms":23435,"temperature":1.0,"reasoning_tokens":3418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:27:33.589555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out the true masks for the unlabeled split and measure the correlation between Eq. (6)'s teacher-student IoU and the true IoU of the teacher pseudo-label; if that correlation is near zero or negative, the sampling and weighting cannot be the mechanism behind the reported 8.34-point gain.","supporting_citations":[{"cited_title":"Scanrefer: 3d object localization in rgb-d scans using natural language,","cited_arxiv_id":null,"evidence_quote":"Supplies the ScanRefer benchmark, evaluation metric, and the labeled/unlabeled splits used for the 1%, 2%, 5%, and 10% regimes."},{"cited_title":"3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the base 3D-RES model shared by all compared methods, including the supervised loss terms."},{"cited_title":"Refteacher: A strong baseline for semi-supervised referring expression comprehension,","cited_arxiv_id":null,"evidence_quote":"The 2D referring semi-supervised baseline that contributes the burn-in stage, EMA teacher update, and attention-alignment loss used in 3DResT."},{"cited_title":"Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learn- ing results,","cited_arxiv_id":null,"evidence_quote":"Underpins the EMA teacher update that prevents the teacher from overfitting to limited labeled data."}],"review_version":1}