{"id":"ac1299ea-0e19-4241-8216-51bbf3b9be0a","arxiv_id":"2502.04111","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Point-level ambiguity from local label disagreement is used to shrink contrastive margins, even to negative values, and this improves 3D semantic segmentation on S3DIS and ScanNet.","lead":"This paper adds a new training rule for 3D point cloud segmentation that relaxes the loss at ambiguous boundary points. It reports consistent accuracy gains over a strong baseline on two indoor scene benchmarks, S3DIS and ScanNet.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ambiguity estimator is unvalidated; if label-disagreement proxy does not track feature-space ambiguity, adaptive margins are random reweighting and the central claim is unsupported.","rationale":"The reader's strongest claim correctly identifies the adaptive margin objective as the central assertion, and the weakest assumption is exactly the unvalidated mapping from position-space label disagreement to per-point ambiguity. My stress-test agrees: the paper's only evidence for this mapping is the ablation in Table III, which varies the margin parameters but never varies or validates the ambiguity signal itself. The proposed shuffle test directly isolates whether the specific assignment of ai matters; if shuffling does not hurt performance, then the method's success cannot be attributed to ambiguity-awareness, and the central claim would fail. The correlation check with prediction entropy provides a complementary mechanistic test. Since the reader already issued a CONDITIONAL verdict based on missing robustness checks, my concern does not change that verdict; it sharpens the key condition: the ambiguity proxy must be validated before the improvement can be attributed to the method's stated motivation. Therefore, the verdict remains CONDITIONAL, with the additional explicit requirement to test the proxy.","tokens_in":9141,"tokens_out":16279,"duration_ms":173248,"concrete_test":"Retrain on S3DIS Area 5 with the same hyperparameters but replace each point's ambiguity ai with a randomly permuted value from the original set (preserving the distribution and thus the margin distribution). If the mIoU does not drop toward the 70.5 baseline, the position-based proxy is not load-bearing and the ambiguity-awareness claim is unsupported. As a secondary check, compute the Spearman correlation between ai and the entropy of the trained model's softmax predictions; low correlation would indicate the proxy does not track feature-space ambiguity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire mechanism of AMContrast3D rests on the ambiguity estimate ai defined in Section III-B (Eqs. 1-4): ai is computed solely from ground-truth label disagreement among K Euclidean nearest neighbors in position space, then converted to a margin mi = μ·ai + ν (Eq. 7) in the contrastive objective (Eqs. 9-11). The paper never validates that this position-space, label-based proxy actually measures the feature-space ambiguity that motivates relaxing constraints. Table III ablates only the margin parameters μ and ν, not the proxy itself; it does not test whether the specific assignment of ai is necessary for the observed gains. If the proxy is incorrect or miscalibrated, the method reduces to a per-point reweighting of the contrastive loss that could improve mIoU for reasons unrelated to ambiguity-awareness. The paper's own argument that labels near transition regions are unreliable further weakens the proxy: ai is derived from exactly those potentially noisy labels, so the margin may be set by annotation noise rather than by a meaningful measure of learning difficulty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AMContrast3D, a supervised contrastive learning method for 3D point cloud semantic segmentation with per-point adaptive margins. The method first estimates a scalar ambiguity per point from the label agreement of its K nearest neighbors in position space (Eqs. 1-4), then maps the ambiguity to a margin mi = mu*ai + nu (Eq. 7). The margin is inserted into the supervised contrastive objective (Eqs. 9-10), which is jointly trained with cross-entropy (Eq. 11). Experiments on S3DIS Area 5 and ScanNet report mIoU gains over the PointNeXt baseline of about 1.3 and 1.4 points, and an ablation on S3DIS shows that the best configuration uses negative margins for high-ambiguity points.","tokens_in":9422,"tokens_out":7412,"duration_ms":77730,"significance":"If the central mechanism is sound, the paper makes a modest but useful contribution: it introduces a simple, parameter-light way to make contrastive objectives adaptive to per-point difficulty in 3D segmentation, and the improvements are consistent across two widely used benchmarks. The formulation is internally coherent, and the ablation supports the claim that negative margins, in particular, are beneficial. The main strengths are the clean integration of a margin generator into an existing backbone and the explicit ablation of the margin mapping. The principal weakness is that the ambiguity estimator itself is not validated, so the mechanistic interpretation of the gains remains open. The significance is therefore conditional: the method is plausible and potentially reproducible, but its conceptual claim requires additional evidence.","major_comments":[{"comment":"The ambiguity proxy ai is the sole driver of the adaptive margin, yet the paper never validates that position-space label disagreement actually measures the feature-space ambiguity that motivates the method. The ablation in Table III varies only the margin parameters mu and nu; it does not test whether the specific assignment of ai carries the signal. We request an experiment that breaks the link between ai and the labels/geometry while keeping the same marginal distribution, for example by permuting ai across points, or by using a constant or reversed assignment. In addition, report on a held-out set the correlation of ai with per-point prediction error or model confidence. If the proxy mainly tracks annotation noise in transition regions, as the paper itself argues those labels are unreliable, the adaptive margins may be set by noise rather than by meaningful ambiguity.","section":"Section III-B, Eqs. (1)-(4) and Table III"},{"comment":"All reported numbers are single runs, and the improvements over the PointNeXt baseline are 1.3 and 1.4 mIoU. For 3D semantic segmentation, these differences are small enough that run-to-run variance could change the ranking. Please report the mean and standard deviation over at least three random seeds for the main comparison, and if possible a paired significance test. This is necessary to support the claim that the improvement is not an artifact of a single run.","section":"Tables I and II"},{"comment":"The margin parameters are set per dataset (mu=-1, nu=0.5 for S3DIS; mu=-1, nu=0.6 for ScanNet) but the paper gives no sensitivity analysis or protocol for selecting them. The ablation in Table III covers only five manually chosen settings on S3DIS, and the ScanNet value of nu=0.6 is not justified. Please provide a sensitivity sweep over nu (and, secondarily, mu and tau) and state how the final values were chosen, for example by a validation split. Otherwise it is unclear whether the reported gain comes from the adaptive mechanism or from per-dataset hyperparameter tuning.","section":"Section IV-A and IV-C"},{"comment":"The ambiguity estimate is computed from ground-truth labels of K nearest neighbors, which are exactly the labels that the paper argues are unreliable in transition regions. This raises a correctness risk: if label noise is concentrated at the same transition points, ai may encode noise rather than task difficulty, and the margin reweighting may be responding to that noise. We recommend a robustness experiment in which the training labels are synthetically perturbed near boundaries, and checking whether the method's advantage (and the ordering of margins) remains stable. This is a direct test of the mechanism proposed in Eqs. (7)-(10).","section":"Section III-B and III-C"}],"minor_comments":[{"comment":"The quantities cc+ and cc- are called closeness centrality, but the formula is the inverse of the mean squared distance, not the standard closeness centrality over shortest paths. Please align the terminology with the cited definition or use a different name such as 'inverse mean squared distance'.","section":"Section III-B, Eqs. (1)-(2)"},{"comment":"The function G is described as an 'inverse sigmoid,' but it is in fact a decreasing logistic (sigmoid) function. The wording should be corrected to avoid confusion.","section":"Section III-B, Eq. (3)"},{"comment":"The implementation details omit several items needed for reproduction: the optimizer and its settings, batch size, weight decay, and the label-embedding strategy borrowed from CBL. Please specify these.","section":"Section IV-A"},{"comment":"On ScanNet validation, AMContrast3D achieves 72.5 mIoU, which is lower than PointMetaBase's 72.8, yet the text says the method 'outperforms' without qualifying that this refers to the test set. Please clarify the comparison.","section":"Table II and Section IV-B"},{"comment":"The phrase 'significant improvements' is used without a statistical test. Given the single-run results, please replace it with 'reported improvements' or add significance testing.","section":"Section IV-C"},{"comment":"The notation DB+ and DB- is used both for decision boundaries and for the inequalities defining the margin. Please separate the boundary plane from the margin region to avoid ambiguity.","section":"Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core idea is interesting, but the central claim depends on an unvalidated ambiguity proxy and on single-run experiments with modest gains. The requested additions—proxy validation, error bars, and sensitivity analysis—are standard and would determine whether the mechanism is as advertised. No concerns about novelty disclosure beyond the provisional arXiv posting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on 3D segmentation. The paper does something simple and sensible: estimate per-point ambiguity from label disagreement in position space, then use that to set a per-point margin in a supervised contrastive loss, with negative margins allowed for very ambiguous points. The loss is coherent, the implementation is straightforward, and the reported gains over PointNeXt are small but consistent on S3DIS and ScanNet. The ablation supports the negative-margin component, which is the least obvious design choice.\n\nThe soft spots are real. The ambiguity proxy is never validated. The paper asserts that label disagreement among K Euclidean neighbors measures the ambiguity that matters for feature learning, but it does not show that this proxy tracks anything meaningful in feature space, nor does it compare against alternative proxies or a random reweighting control. If the proxy is just noise, the method reduces to a per-point reweighting that works for some other reason, and the paper's story is wrong even if the numbers hold. A reviewer should ask for a sanity check here.\n\nThe experiments are also single-run, with no error bars, and the margin hyperparameters mu and nu are tuned per dataset. The gains are modest, so without variance estimates it is hard to know if 1.3-1.4 mIoU is meaningful. No code is released.\n\nThat said, the paper is honest: it claims modest improvements, the method is backbone-agnostic, and it does not oversell. The central empirical claim is plausible. The citation pattern looks fine. This is not a derivation paper, so the circularity burden is low; it is an empirical method paper.\n\nWho is it for? Researchers working on point cloud segmentation or margin-based losses. It deserves a serious referee: the idea is novel enough, and the experiments are adequate for a solid conference paper pending robustness checks. I would recommend the editor send it to review and ask for error bars, a control with a random ambiguity proxy, and ideally code or feature-space validation. If those checks hold, this is a useful contribution.","headline":"A sensible, modestly effective 3D segmentation method whose ambiguity proxy is plausible but unproven; worth a serious look but needs robustness checks.","tokens_in":9902,"tokens_out":2062,"would_cite":false,"duration_ms":21227,"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":"The paper claims that per-point ambiguity—estimated from label disagreement among a point's nearest neighbours—should set the margin of a supervised contrastive loss, with negative margins for highly ambiguous points, and that this lifts…","keywords":["3D semantic segmentation","point cloud","contrastive learning","adaptive margin","ambiguity estimation","transition regions","supervised contrastive loss"],"falsifier":"Train the same PointNeXt architecture with the AMContrast3D loss on S3DIS Area 5, but replace the estimated ambiguity $a_i$ with random per-point margins drawn from the same distribution of $m_i$; if random margins match the reported mIoU, the improvement is not driven by the ambiguity signal. A complementary check is to corrupt the neighbour labels used by the estimator while keeping the training labels intact: if performance does not drop, the estimator is not the source of the gain.","tokens_in":8969,"feed_emoji":"🎯","tokens_out":5804,"duration_ms":53445,"temperature":0.7,"pith_summary":"AMContrast3D argues that 3D semantic segmentation fails when the training loss treats every point equally, because points near transitions between classes are intrinsically ambiguous and their labels are unreliable. The paper's proposal is to read each point's ambiguity from the labels of its Euclidean nearest neighbors and convert that ambiguity into a per-point margin in a supervised contrastive loss: low-ambiguity points get positive margins that force hard separation, borderline points get zero margins, and highly ambiguous points get negative margins that relax the objective. With this schedule, the model should concentrate learning on points it can actually get right while not being derailed by mislabeled or undecidable points. On S3DIS Area 5 and ScanNet, the method reports higher mIoU than the PointNeXt baseline and than prior contrastive-boundary methods. If correct, it would mean that decision boundaries in dense 3D prediction can be usefully per-point and data-dependent rather than uniform.","feed_headline":"Adaptive margins lift 3D segmentation on ambiguous points","feed_subtitle":"Label disagreement among nearby points shrinks or reverses contrastive margins, raising mIoU on S3DIS and ScanNet.","key_machinery":"The load-bearing object is the pair formed by the ambiguity estimator and the margin generator. The estimator uses position embeddings $p_i$ to separate the $K$-nearest neighbourhood into intra-points $N^+_i$ (same label) and inter-points $N^-_i$ (different label), computes closeness centralities $cc^+_i = |N^+_i|/d^+_i$ and $cc^-_i = |N^-_i|/d^-_i$, and maps their difference through an inverse sigmoid (with piecewise endpoints at 0 and 1) to get $a_i$. The margin generator then sets $m_i = \\mu a_i + \\nu$, and the contrastive objective in Eq. (10) uses $\\exp((\\mathrm{sim}(f_i,f_j)-m_i)/\\tau)$ for intra-pairs, so the effective required separation between intra- and inter-similarity is $m_i$ instead of 0. With $\\mu=-1$ and $\\nu=0.5$ on S3DIS or $\\nu=0.6$ on ScanNet, low-ambiguity points face a positive margin, semi-ambiguous points face zero, and the most ambiguous points face a negative margin, which is what lets the training difficulty vary point by point.","core_discovery":"The central claim is that the additive margin idea from classification can be transplanted into point-level supervised contrastive learning for 3D point clouds, with the margin made a linear function of an estimated per-point ambiguity: $m_i = \\mu a_i + \\nu$. The ambiguity $a_i \\in [0,1]$ is computed from position embeddings by counting how many of the point's $K$ nearest neighbours disagree with its label and comparing their closeness centralities; points surrounded only by same-label neighbours get $a_i=0$, and points whose neighbourhood is fully conflicting get $a_i=1$. Injecting this margin into the exponent of the contrastive softmax, $\\exp((\\mathrm{sim}(f_i,f_j)-m_i)/\\tau)$, shifts the decision boundary between intra-class and inter-class similarity so that the required separation shrinks or reverses as ambiguity grows. The paper reports that including this adaptive term, weighted by $1-\\lambda$ alongside cross-entropy, improves mIoU to 71.8% on S3DIS Area 5 and 72.6% on ScanNet test, and its ablation shows the gain disappears when the margin is constant or clipped to be non-negative.","pith_inferences":["The same ambiguity-to-margin map could be read as a self-derived label-noise weighting: it downweights exactly the points whose neighbour disagreement makes labels suspect, so the mechanism may be a general ambiguity-weighting principle rather than a specifically contrastive one.","One testable extension is to feed the estimated ambiguity $a_i$ into the cross-entropy term as well, for instance as instance-dependent label smoothing or loss weights; if the gains persist, the margin is a vehicle for a broader ambiguity-aware objective.","If the ambiguity proxy is sound, it could transfer to other dense 3D tasks such as instance segmentation or object detection, where transition regions between objects are also the hardest to annotate consistently.","A stronger test would replace the inverse-sigmoid curve with the raw count of disagreeing neighbours; the paper's ablations do not isolate the centrality weighting from the count, so the contribution of the closeness term remains open."],"forward_implications":["Training difficulty becomes a per-point quantity: the same loss formula applies at every point but with decision boundaries that depend on the local label configuration of the scene.","Points sitting on semantic boundaries are explicitly de-emphasized, so gradients concentrate on interior points whose labels are reliable, which should make training more stable than a uniform contrastive term.","The margin generator is decoupled from the backbone and can be attached to any point-level contrastive loss used in supervised 3D segmentation.","Negative margins are reported as essential: clipping them at zero (the last row of the ablation) drops mIoU from 71.8% to 70.5%, indicating that the relaxation, not merely the adaptivity, carries part of the gain."],"supporting_citations":[{"why":"Supplies the encoder-decoder backbone and the baseline whose mIoU the method improves.","marker":"[4]"},{"why":"Provides the supervised contrastive formulation and label-mining strategy for point cloud segmentation that AMContrast3D extends with margins.","marker":"[7]"},{"why":"Defines the temperature-scaled supervised contrastive objective whose decision boundaries the margins modify.","marker":"[27]"},{"why":"Supplies the S3DIS benchmark used for the Area 5 evaluation.","marker":"[33]"},{"why":"Supplies the ScanNet benchmark used for the validation and test evaluation.","marker":"[34]"},{"why":"Supplies the additive angular margin idea from face recognition that motivates margin-based decision boundaries.","marker":"[10]"},{"why":"Provides a second margin-loss source justifying the shifting of decision boundaries in the contrastive softmax.","marker":"[11]"},{"why":"Gives the closeness-centrality notion used in the ambiguity estimator's $cc^+$ and $cc^-$ terms.","marker":"[29]"}],"fun_headline_variants":["Adaptive margins boost 3D segmentation on ambiguous points","Per-point ambiguity drives contrastive margins in 3D","Shrinking margins for ambiguous points lift 3D segmentation","Adaptive contrastive margins conquer ambiguous 3D points","Ambiguity-aware margins improve 3D point cloud segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the premise that a point's ambiguity—and therefore the right amount of training pressure—can be read off from label disagreement among its nearest neighbours in 3D position space, even though the paper itself notes that labels near transition regions are questionable for human annotators.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive margins boost 3D segmentation on ambiguous points","Per-point ambiguity drives contrastive margins in 3D","Shrinking margins for ambiguous points lift 3D segmentation","Adaptive contrastive margins conquer ambiguous 3D points","Ambiguity-aware margins improve 3D point cloud segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1452,"prompt_tokens":966,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":403}},"tokens_in":582,"tokens_out":486,"duration_ms":4496,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:27:21.222739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same PointNeXt architecture with the AMContrast3D loss on S3DIS Area 5, but replace the estimated ambiguity $a_i$ with random per-point margins drawn from the same distribution of $m_i$; if random margins match the reported mIoU, the improvement is not driven by the ambiguity signal. A complementary check is to corrupt the neighbour labels used by the estimator while keeping the training labels intact: if performance does not drop, the estimator is not the source of the gain.","supporting_citations":[{"cited_title":"23192–23204, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the encoder-decoder backbone and the baseline whose mIoU the method improves."},{"cited_title":"8489–8499","cited_arxiv_id":null,"evidence_quote":"Provides the supervised contrastive formulation and label-mining strategy for point cloud segmentation that AMContrast3D extends with margins."},{"cited_title":"18661–18673, 2020","cited_arxiv_id":null,"evidence_quote":"Defines the temperature-scaled supervised contrastive objective whose decision boundaries the margins modify."},{"cited_title":"Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner, ”ScanNet: Richly-annotated 3D reconstructions of indoor scenes,” in CVPR, 2017, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the ScanNet benchmark used for the validation and test evaluation."},{"cited_title":"4690–4699","cited_arxiv_id":null,"evidence_quote":"Supplies the additive angular margin idea from face recognition that motivates margin-based decision boundaries."},{"cited_title":"5265–5274","cited_arxiv_id":null,"evidence_quote":"Provides a second margin-loss source justifying the shifting of decision boundaries in the contrastive softmax."},{"cited_title":"Prokopyev, and Eduardo L","cited_arxiv_id":null,"evidence_quote":"Gives the closeness-centrality notion used in the ambiguity estimator's $cc^+$ and $cc^-$ terms."}],"review_version":1}