{"id":"2a477208-21df-4ee8-8593-05ac32264f13","arxiv_id":"2508.09626","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"SAD-Splat combines semantic-aware pruning of 3D Gaussian points with pseudo-label supervision from 2D foundation models to improve aerial-view semantic segmentation accuracy and representation compactness.","lead":"This preprint proposes SAD-Splat, a method for 3D aerial-view semantic segmentation that prunes redundant 3D Gaussian points using learned semantic confidence and trains with pseudo-labels from 2D foundation models when ground-truth labels are sparse. It also introduces a new benchmark dataset, 3D-AS, for aerial scenes with sparse annotations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy/compactness balance hinges on an unvalidated confidence signal: low-confidence pruning trained with high-confidence pseudo-labels can discard informative points and inflate apparent accuracy, so the central claim needs a dense-label ablation.","rationale":"The reader identified the same load-bearing assumption: semantic confidence scores must reliably identify ambiguous and redundant Gaussian points across scale and occlusion, and pseudo-labels can reinforce the model's own errors. I see no reason to disagree. On a good-faith reading, the method is coherent: Hard Concrete sparsification is a known technique, pseudo-labeling is a standard strategy for limited supervision, and a new aerial benchmark could be a useful community resource. The concern is not that the method disagrees with consensus; it is that the two novel components share a single unvalidated confidence signal, so both the compactness gain and the maintained accuracy could be driven by self-training dynamics rather than by genuine identification of redundant or ambiguous points. This is directly testable with dense labels, per-class metrics, calibration diagnostics, and a random-pruning control. Because the reader's verdict was already CONDITIONAL and these tests fit naturally as conditions for acceptance, the verdict remains UNCHANGED rather than moving to a different category. The abstract-level claim should not be treated as established until the confidence signal is validated against dense ground truth and shown to outperform random pruning at equal compression.","tokens_in":22851,"tokens_out":4502,"duration_ms":51018,"concrete_test":"On held-out test scenes, obtain dense manual semantic labels and compute: (i) per-class IoU for small and occluded classes with and without the drop module at matched point budgets; (ii) expected calibration error or reliability diagrams for the semantic confidence used by the drop module, evaluated against those dense labels; and (iii) a control with random Gaussian pruning at identical compression ratios. If random pruning matches or beats SAD-Splat, or if calibration error is high on the classes that are pruned most, then semantic-confidence pruning is not the source of the reported accuracy/compactness balance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the accuracy/compactness balance, and the specific mechanism claimed to produce it is the Gaussian point drop module: points with low semantic confidence are stochastically masked via a Hard Concrete gate. The load-bearing assumption is that this confidence signal measures semantic ambiguity or redundancy in aerial scenes. That assumption is not established, and the paper's own design creates a specific failure mode. The same paper trains the confidence-producing segmentation head with high-confidence pseudo-labels from 2D foundation models when ground-truth labels are sparse. This couples the two components: a Gaussian whose prediction disagrees with the teacher receives low confidence, and low confidence is exactly the criterion for pruning. In a self-training loop, the model can become more self-consistent without becoming more correct, so pruned points may be the ones the model has learned to ignore rather than truly redundant ones. At aerial scale, small objects and occlusion boundaries occupy few Gaussians, so per-point confidence is least reliable exactly where pruning is most damaging. The garbled full text provides no inspectable dense-label, per-class, or calibration numbers, so the claimed balance is currently unverified rather than demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAD-Splat, a 3D aerial-view semantic segmentation method based on 3D Gaussian Splatting. It introduces a Gaussian point drop module that combines semantic confidence estimation with a Hard Concrete sparsity mechanism to prune redundant or ambiguous points, and a high-confidence pseudo-label generation pipeline that uses 2D foundation models to supervise training when ground-truth labels are sparse. The abstract also announces a new benchmark, 3D-AS, and claims that the method achieves an excellent balance between segmentation accuracy and representation compactness. However, the submitted full text is almost entirely corrupted and unreadable: equations, tables, figures, and most prose cannot be inspected, so the technical content and experimental support for these claims are not available for evaluation.","tokens_in":23189,"tokens_out":3672,"duration_ms":36810,"significance":"If the claims were substantiated, the idea of confidence-guided pruning with Hard Concrete gates for 3D aerial segmentation would be an interesting contribution, potentially reducing storage and computation while preserving segmentation quality. The high-confidence pseudo-label pipeline addresses the practically important issue of sparse annotations in aerial scenes, and a new benchmark such as 3D-AS could be useful to the community. The paper, however, provides no inspectable quantitative evidence: no numbers, error bars, baseline comparisons, dataset statistics, or implementation details are visible. The strengths are asserted only in the abstract, and the method's central mechanism, confidence-based pruning, is not validated in the readable portions of the manuscript.","major_comments":[{"comment":"The body of the manuscript is almost entirely garbled and unreadable; equations, tables, figures, and most prose cannot be inspected, so the central claim of an 'excellent balance' between accuracy and compactness is unsupported by any visible numbers. This blocks verification of every load-bearing result, including the pruning ablation, the pseudo-label gains, and the 3D-AS benchmark statistics, and the paper cannot be accepted in this form.","section":"Full text (all sections after Abstract)"},{"comment":"The pruning mechanism relies on per-point semantic confidence as the signal for redundancy, but the readable text provides no calibration analysis, no per-class results, and no dense-label ablation. Because the same confidence signal is used to select high-confidence pseudo-labels when ground-truth labels are sparse, the paper needs to show that pruning removes truly redundant points rather than points whose predictions disagree with the teacher; please report per-class IoU with and without the drop module on dense labels, as well as confidence calibration curves.","section":"Abstract / Gaussian point drop module"},{"comment":"No baseline comparisons, dataset statistics, error bars, or implementation details are visible, and the Hard Concrete sparsity target and temperature are free parameters that directly control the compactness/accuracy trade-off. The manuscript must report a sensitivity analysis over these parameters, together with dense-label and pseudo-label variants, to substantiate the claimed balance between segmentation accuracy and representation compactness.","section":"Experiments (unreadable)"}],"minor_comments":[{"comment":"The abstract uses both '3D-AVS-SS' and '3D-AS' inconsistently; please make the acronym usage consistent throughout the paper.","section":"Title / Abstract"},{"comment":"The full text includes the line 'arXiv:2508.09627v1 [cs.LG] 13 Aug 2025', which appears to belong to a different paper; this suggests a corrupted or misassembled source file that must be corrected.","section":"Full text header"},{"comment":"The reference list and related-work discussion are not readable in this version; please ensure the bibliography and related-work section are properly rendered so that prior work and novelty can be assessed.","section":"References and related work"}],"recommendation":"uncertain","confidential_remarks":"The manuscript cannot be technically reviewed in its current form because the full text is corrupted and unreadable. I recommend asking the authors to resubmit a clean PDF containing the complete text, tables, figures, and references before an actual technical review. If the corruption is due to a submission or conversion pipeline issue, please disregard this report and request a corrected submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a coherent-sounding systems paper whose abstract claims a strong accuracy/compactness trade-off, but the supplied full text is garbled and no quantitative results are visible. You can judge the idea, not the execution.\n\nWhat is new: the specific combination of Hard Concrete sparsity with semantic-confidence-based pruning for aerial 3D Gaussian splatting, plus a pseudo-label pipeline from 2D foundation models for sparse-annotation settings. The proposed 3D-AS benchmark could be a useful community resource if it is real and released. The abstract is well organized and the method has a clear internal logic.\n\nSoft spots, in proportion: the central claim is asserted, not demonstrated. There are no numbers, no baseline comparisons, no error bars, no dataset statistics, and no code or data. That is a real gap for any empirical paper. The stress-test note raises a genuine concern: the pruning criterion is low semantic confidence, while the training signal comes from high-confidence pseudo-labels. If low confidence mostly means \"disagrees with the teacher,\" pruning could remove informative points and the model could become more self-consistent without becoming more correct. That is a plausible failure mode, but it is a hypothesis, not a demonstrated flaw. A dense-label ablation and per-class results would settle it. I would not call the paper circular or incoherent; it is just unverified.\n\nThe math fragments in the garbled full text suggest the authors have actually worked through the Hard Concrete gate and the loss terms, which is more than many abstracts of this type. The citation pattern is not something I can judge from the text provided, and self-citation is not an issue on its own.\n\nWho this is for: people working on 3D Gaussian splatting, aerial segmentation, or model compression for scene understanding. They would want to see the full version with experiments. As it stands, the paper is a plausible contribution that deserves a serious referee rather than a desk reject. My recommendation: send it to review, but make clear that acceptance should hinge on released code/data and dense-label ablation evidence that the pruning signal is not just self-consistency.","headline":"Plausible system paper with a useful dataset idea, but the central accuracy/compactness claim is unverified because no results or code are visible; worth sending to review, not desk-rejecting.","tokens_in":23594,"tokens_out":1756,"would_cite":false,"duration_ms":18836,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SAD-Splat prunes redundant and ambiguous 3D Gaussian points in aerial scenes using semantic confidence and a learnable Hard Concrete sparsity mask, while pseudo-labels from 2D foundation models handle sparse ground truth.","keywords":["3D aerial-view semantic segmentation","Gaussian splatting","point pruning","Hard Concrete distribution","semantic confidence","pseudo-labels","sparse annotations","representation compactness"],"falsifier":"Take a held-out portion of an aerial scene with dense ground-truth labels and compare SAD-Splat's dropped Gaussian points against those labels. If a large fraction of dropped points lies on object boundaries, small structures, or regions the model misclassifies, or if perturbing the confidence threshold sharply changes segmentation accuracy, the claim that ambiguous points are being removed would be contradicted. A second check is to train with pseudo-labels disabled and compare with the full pipeline; if pseudo-labels do not improve accuracy under sparse labels, the supervision claim fails.","tokens_in":22653,"feed_emoji":"🛰️","tokens_out":4352,"duration_ms":39900,"temperature":0.7,"pith_summary":"The paper tries to establish that semantic segmentation of 3D aerial scenes can be made both more accurate and more compact by treating redundant Gaussian points as a removable source of ambiguity. It claims that a Gaussian point drop module, scoring each point by semantic confidence and letting a learnable Hard Concrete distribution decide which points to zero out, removes exactly the points that hurt segmentation while shrinking the representation. It further claims that high-confidence pseudo-labels generated by 2D foundation models can substitute for missing ground-truth annotations in aerial imagery. If right, this gives a scalable path to segmenting large aerial scenes from sparse labels without sacrificing quality, and it provides a new benchmark, 3D-AS, to measure progress. The sympathetic reading is that pruning is not a compression afterthought but a way to force the model to commit to confident semantic decisions.","feed_headline":"A pruning rule keeps 3D aerial scenes accurate with fewer points","feed_subtitle":"Confidence-aware point dropping plus pseudo-labels keep maps lean without losing segmentation quality.","key_machinery":"The central object is the Gaussian point drop module. It works by first assigning each 3D Gaussian a semantic confidence score, then applying a learnable sparsity mask sampled from the Hard Concrete distribution, which is a continuous relaxation of discrete Bernoulli masks that keeps the pruning decision differentiable. The mask learns to eliminate points that are redundant or semantically ambiguous, so the surviving Gaussians carry clearer semantic signal. The second mechanism is a high-confidence pseudo-label pipeline in which a 2D foundation model proposes labels for unlabeled areas and only high-confidence predictions enter training. Together these form the argument: pruning removes confusion, and pseudo-labels supply supervision that sparse real-world aerial annotations lack.","core_discovery":"The paper's central discovery claim is that redundant and semantically ambiguous Gaussian points are not merely wasteful; they actively degrade 3D aerial-view semantic segmentation. SAD-Splat therefore couples semantic confidence estimation with a learnable sparsity mechanism based on the Hard Concrete distribution, so that the model itself learns which points to drop. The same pipeline also generates high-confidence pseudo-labels from 2D foundation models to strengthen supervision when ground-truth labels are sparse. On the new 3D-AS benchmark, the authors report that SAD-Splat keeps segmentation accuracy while improving representation compactness, offering an efficient and scalable solution for 3D aerial scene understanding.","pith_inferences":["Editorial inference: the confidence scores from 2D foundation models are themselves predictions, so if they are systematically overconfident for aerial views, pruning and pseudo-labeling could jointly reinforce errors; a held-out dense-label comparison would reveal this.","Editorial inference: the Hard Concrete sparsity mechanism is not specific to aerial Gaussians, so the same pruning principle could transfer to other 3D representations or to tasks like object detection, wherever redundant points create ambiguity.","Editorial inference: the pruning step may act as an indirect regularizer, and one testable question is whether it improves accuracy on fully labeled scenes or only in the sparse-label regime where pseudo-labels dominate."],"forward_implications":["Deploying SAD-Splat on large aerial scenes would require less memory and faster rendering because redundant Gaussians are removed during training.","Segmentation consistency across scale variations and occlusions should improve because ambiguous points are pruned rather than classified with low confidence.","Sparse-annotation aerial datasets become more usable through high-confidence pseudo-labels, reducing the manual labeling cost for 3D aerial scene understanding.","The new 3D-AS benchmark gives future aerial-view segmentation methods a shared evaluation target with a focus on real-world sparse annotations."],"supporting_citations":[],"fun_headline_variants":["Prune redundant Gaussians, sharpen 3D aerial segmentation","SAD-Splat: learnable sparsity for accurate 3D aerial scene parsing","Confidence-based Gaussian dropping keeps 3D aerial maps lean","Hard Concrete sparsity trims Gaussians for better 3D aerial views","Adaptive pruning of ambiguous points improves 3D aerial segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the semantic confidence scores produced by 2D foundation models, when applied to rendered 3D aerial Gaussians, correctly identify which points are redundant or ambiguous; if those scores are miscalibrated for aerial imagery, pruning will remove informative points and pseudo-labels will reinforce the model's own mistakes.","fun_headline_variants_meta":{"raw":{"variants":["Prune redundant Gaussians, sharpen 3D aerial segmentation","SAD-Splat: learnable sparsity for accurate 3D aerial scene parsing","Confidence-based Gaussian dropping keeps 3D aerial maps lean","Hard Concrete sparsity trims Gaussians for better 3D aerial views","Adaptive pruning of ambiguous points improves 3D aerial segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1542,"prompt_tokens":880,"completion_tokens":662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":496,"tokens_out":662,"duration_ms":6135,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:54:20.336348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out portion of an aerial scene with dense ground-truth labels and compare SAD-Splat's dropped Gaussian points against those labels. If a large fraction of dropped points lies on object boundaries, small structures, or regions the model misclassifies, or if perturbing the confidence threshold sharply changes segmentation accuracy, the claim that ambiguous points are being removed would be contradicted. A second check is to train with pseudo-labels disabled and compare with the full pipeline; if pseudo-labels do not improve accuracy under sparse labels, the supervision claim fails.","supporting_citations":[],"review_version":1}