{"id":"4706e893-f578-4a9c-b826-8f5cbc4d9bd7","arxiv_id":"2504.19581","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SAMBLE learns shape-specific point sampling strategies using a sparse attention map and per-bin sampling weights to balance local detail and global uniformity.","lead":"This paper introduces SAMBLE, a learning-based method that samples points from 3D point clouds while balancing sharp edge details and overall shape coverage. It reports improved accuracy on standard 3D classification and segmentation benchmarks, including when only a few points are kept.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains over APES rest on test-set-selected hyperparameters and an unreported k; without code or a validation split, the central performance claim is not yet independently established.","rationale":"The reader's weakest_assumption is the density confound of the kNN asymmetry frequency. That is a plausible and important limitation, but it primarily affects generalization to real-world non-uniform scans, which the paper explicitly leaves to future work ('adapting the proposed method for point cloud scenes is another promising area to explore'). The more load-bearing issue for the central claim as actually evaluated is that the reported benchmark numbers are not established by a sound evaluation protocol: hyperparameters appear to be selected on the test set, k is unreported, no code is released, and no variance is given. These are all facts visible in the manuscript rather than outside-consensus assumptions. The concern does not require rejecting the method; it requires conditional acceptance pending a validation-based re-run and code release. The reader already reached CONDITIONAL, so the verdict is unchanged, but the emphasis shifts from the density concern to the reproducibility and selection-on-test issue.","tokens_in":21602,"tokens_out":14666,"duration_ms":170958,"concrete_test":"Release code with the missing k specified, and re-run the main classification (Table 2) and few-point (Table 4) experiments using a proper train/validation/test split. Select n_b, gamma, tau, and the indexing mode only on the validation split, then evaluate that single configuration on the test set and report mean and standard deviation over at least three seeds. If the validation-chosen accuracy is not above APES (global and with pre-processing) by at least the reported margins, the central empirical claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: SAMBLE is superior to APES and other baselines on classification, segmentation, and few-point sampling. The evidence is the test-set accuracy in Tables 2-4. However, the design hyperparameters appear to have been chosen by inspecting exactly that test benchmark: the indexing mode (Table 5), the number of bins (Table 6), the momentum factor (Table 11), and the temperature (Table 12) are all ablated on the test set, and the best configurations are then used in the main results. No held-out validation split is described, no code is released, and the k in the k-NN sparse attention map (Sec. 3.1, Table 1) is never given a numeric value anywhere in the paper or supplement. The margins over APES in Table 2 are small (0.3-0.4 accuracy points), and the experiments report a single run without seed variance. The method may be sound, but the load-bearing numerical comparison is not independently checkable as reported: the reported 'superior performance' could be a selection artifact rather than a property of the sampling strategy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAMBLE, a learning-based point cloud sampling method that computes point-wise sampling scores from a sparse attention map (SAM) combining local and global information, and then samples points using a shape-specific bin-based strategy. The sparse attention map is built by masking a global attention map with kNN neighbor indices, and seven indexing modes are proposed for deriving sampling scores. Points are partitioned into bins with momentum-updated boundaries, and bin sampling weights are learned with additional bin tokens. The method is evaluated on ModelNet40 classification, ShapeNet-Part segmentation, few-point sampling on ModelNet40, and intermediate sub-point cloud segmentation, reporting state-of-the-art or competitive results against APES and other baselines. The paper includes extensive ablations over indexing modes, number of bins, momentum factor, temperature, and upsampling layers, plus supplementary material on network architecture, training details, and additional visualizations.","tokens_in":21821,"tokens_out":3980,"duration_ms":43141,"significance":"If the reported results are reliable, SAMBLE makes a useful contribution to task-oriented point cloud sampling by offering a more balanced local-global trade-off than APES, with particularly strong few-point sampling performance. The paper's systematic exploration of multiple indexing modes, the bin-based shape-specific sampling mechanism, the qualitative score heatmaps, and the runtime/complexity analysis are strengths. The central claim is empirical, however, and its validity depends on the evaluation methodology; the selection of design hyperparameters on the test benchmarks and the lack of statistical variance reporting are serious concerns. The method is also not independently reproducible as reported because the kNN neighborhood size k is never specified and no code is provided.","major_comments":[{"comment":"The main results in Tables 2-4 are produced with hyperparameters that appear to be selected on the same test benchmarks used for the headline numbers. Specifically, the indexing mode is chosen from Table 5, the number of bins from Table 6 (and the corresponding main-text sentences), the momentum factor from Table 11 in the supplementary, and the temperature from Table 12 in the supplementary. No held-out validation split is described anywhere; the paper only mentions the official train-test splits for ModelNet40 and ShapeNet-Part. This is a form of test-set selection that can inflate reported accuracy, and it undermines the claim of 'superior performance' over APES, whose margins are only 0.3-0.4 accuracy points in Table 2. The authors should re-run the ablation selection on a validation split and report the final results on the test set only once, or otherwise demonstrate that the selected configuration is not an artifact of test-set tuning.","section":"Sec. 4.4 and supplementary Tables 11-12"},{"comment":"The kNN neighborhood size k is a central free parameter of the proposed sparse attention map and of every indexing mode in Table 1 (e.g., the number of selected cells per row, and the column counts n_o used in modes v-vii). Throughout the paper and the supplementary material, k is never given a numeric value, and no ablation over k is reported. Without this value, the method cannot be reproduced or compared against other sampling methods. This missing specification is load-bearing for the empirical claim and must be added, along with an ablation or at least the chosen value for each experiment.","section":"Sec. 3.1 and Table 1"},{"comment":"Table 3 reports intermediate sub-point cloud segmentation results where SAMBLE's category mIoU increases as the number of points decreases (84.51 at 2048 points, 84.84 at 1024, 85.04 at 512). This is counterintuitive: a smaller sampled subset would generally not be expected to improve accuracy on a downstream task unless the evaluation protocol differs from what is described. The paper states that 'the learned sampled points contribute more to the overall performance' and that 'the upsampling layer cannot fully reconstruct the features of the discarded points,' but this does not explain why fewer points should yield higher mIoU. The authors should clarify the evaluation protocol for the intermediate sub-point clouds and explain this trend, as it currently undermines the interpretability of a key comparison.","section":"Table 3"},{"comment":"All reported accuracy values appear to come from a single run with no seed variance or statistical significance testing. In Table 2, the gap between SAMBLE and APES (global) is 0.4% overall accuracy on ModelNet40 and 0.8% category mIoU on ShapeNet-Part; on a test set of 2,468 shapes, a 0.4% difference corresponds to roughly 10 shapes and may be within run-to-run variation. The few-point sampling gains in Table 4 are large and more convincing, but the classification and segmentation claims should be supported by mean±std over multiple seeds or, at minimum, a statement of the number of runs and the observed variance.","section":"Tables 2 and 4"}],"minor_comments":[{"comment":"The notation in the indexing-mode formulas is inconsistent: 'oth row' and 'oth column' should be 'o-th row' and 'o-th column', and the definition of S_o is only given in prose rather than in the table. Please align the notation between Table 1 and the surrounding text.","section":"Sec. 3.2"},{"comment":"The phrase 'quantum-entangled twin point pair' is informal and not defined; the phenomenon is simply that two points can each be the other's kNN under ties or boundary conditions. Please replace this with a precise geometric description.","section":"Sec. 1, Fig. 2"},{"comment":"The figure caption states scores are normalized to N(0.5,1) for visualization, but no colorbar or explicit scale is provided. Adding a colorbar and explaining what '1.0 (or higher)' means would improve interpretability.","section":"Fig. 5"},{"comment":"The temperature parameter tau is introduced in Eq. (5) for within-bin softmax sampling, and the supplementary extends it to global prior-based sampling in Eq. (10). The relationship between these two uses is clear but should be stated in the main text for completeness.","section":"Sec. 3.3, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental development over APES from the same group, but the core idea of a sparse attention map and shape-specific bin sampling is a reasonable next step. The test-set hyperparameter selection is the most serious issue and may be fixable: the authors could hold out a validation split from the training set for ablations, then report test results once. If the margins over APES shrink or vanish under that protocol, the central claim would need to be substantially revised. The missing k value and the absence of code are additional reproducibility barriers that the editor should weigh. I recommend major revision rather than rejection because the proposed mechanism is plausible and the few-point sampling gains are large."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. SAMBLE is a genuine, well-scoped extension of APES, not a repackaging: the sparse attention map and the bin-token mechanism are new, and the paper shows they buy consistent gains in classification, part segmentation, and few-point sampling. The few-point results are the strongest part—at M=32 they jump from APES's 88.56 (with preprocessing) to 89.45, and from 75.63 without preprocessing to 89.45. That is not a small margin, and it is hard to explain away as a selection artifact.\n\nThe paper does a lot of things right. The ablations are extensive and mostly well designed: they compare seven indexing modes, vary the number of bins, momentum, temperature, and upsampling type. The supplementary is unusually honest—the discussion of why pre-softmax attention maps are needed for bin weights, and why mean-pooling before ReLU beats the reverse, shows real engagement with the mechanics. The qualitative results, including heatmaps and bin histograms, make the shape-specific behavior concrete.\n\nNow the soft spots, in order of seriousness. First, the design hyperparameters—indexing mode, bin count, momentum, temperature—are selected by ablating on the same ModelNet40 test set that produces the headline numbers. No held-out validation split is described. That is a known bias, and for Table 2, where the margins over APES are only 0.3–0.4 points, it undermines the claim that the whole configuration is better, not just the search. Second, the k in the k-NN sparse attention map is never given a numeric value anywhere. That is a concrete reproducibility gap that no reviewer should wave through. Third, there is no code and no seed variance reported, so the 0.3-point differences are not statistically grounded. The comparison with SOTA methods in Table 2 is also not architecture-controlled, though the authors do match APES's base architecture, which is the right comparator.\n\nThe density-confound concern (kNN asymmetry frequency being affected by local density on real scans) is legitimate but not load-bearing here, since the benchmarks are uniform CAD meshes. It does limit the claim to that setting.\n\nBottom line: the central mechanism appears sound and the few-point gains are substantial, but the reported numbers are not yet independently checkable as presented. This deserves a serious referee. I would send it to review, asking for code, a proper validation split, the k value, and variance estimates. A reader working on point cloud sampling will get value from it now; with those fixes, the evidence would be solid.","headline":"Genuine extension of APES with strong few-point results, but test-set-selected hyperparameters and a missing k make the headline numbers hard to check independently.","tokens_in":22376,"tokens_out":2687,"would_cite":true,"duration_ms":26257,"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":"SAMBLE samples point clouds to keep edge detail and global uniformity, and reports gains over prior samplers.","keywords":["point cloud sampling","sparse attention map","shape-specific sampling","bin-based sampling","edge point preservation","global uniformity","few-point sampling","point cloud classification"],"falsifier":"Run SAMBLE on a real scan with visibly non-uniform point density and compare its top-score points against manually labeled true edges; if dense flat regions are scored as high as genuine sharp edges, the kNN-frequency signal is density, not geometry.","tokens_in":21393,"feed_emoji":"🎯","tokens_out":7194,"duration_ms":67942,"temperature":0.7,"pith_summary":"Point cloud sampling decides which points of a 3D shape survive to a smaller set, and most methods either destroy recognizability by generating artificial points or over-collect sharp edges at the expense of the overall shape. SAMBLE is a learning-based sampler that claims to break that trade-off: it reads a per-point 'edge-ness' signal from a sparse attention map and then learns, per shape, how many points to draw from each score bin. The paper reports that this shape-specific strategy improves ModelNet40 classification to 94.2% overall accuracy, ShapeNet-Part segmentation to 84.5% category mIoU, and produces strong few-point sampling results down to 32 points from raw input. The reason a reader should care is that better sampling is a cheap, drop-in improvement for any point-cloud pipeline that uses FPS or random sampling.","feed_headline":"SAMBLE keeps edges and shape uniformity in sampled point clouds","feed_subtitle":"Per-bin learned sampling balances sharp detail with global coverage, boosting classification and few-point accuracy.","key_machinery":"The load-bearing mechanism is the sparse attention map (SAM) plus a bin-based sampling policy. SAM takes an N×N global attention map and, for each row, carves out only the cells corresponding to that point's k nearest neighbors, setting the rest to zero; the per-column count no and the column sums of these carved cells become the raw material for point-wise sampling scores. The preferred indexing mode, sparse column square-divided, divides each column sum by $no^{2}$, which softens the edge over-emphasis seen in row-standard-deviation modes. The bin policy then divides the score range into nb bins using boundaries updated with momentum over the training set, and nb learnable bin tokens, concatenated to keys and values, produce per-bin sampling weights through a masked point-to-token attention sub-map; within each bin, points are drawn by a softmax over scores with temperature τ. This combination is what turns a fixed top-M rule into a shape-specific strategy.","core_discovery":"The central claim is that the asymmetry of k-nearest-neighbor relations — point pi can be among pj's neighbors without the reverse holding — is an easily computed point property that separates edge points, close-to-edge points, and non-edge points, and that a sparse attention map exposes this property. Starting from a global attention map, SAMBLE keeps only the k selected cells per row; because each point is chosen as a neighbor a different number of times, column variables like the selected-cell count no carry the geometric signal. Among seven indexing modes, the 'sparse column square-divided' mode (column sum divided by $no^{2}$) is chosen as the default because it samples edge points without over-emphasizing them. On top of those per-point scores, bin-based sampling partitions each shape's points into bins with momentum-updated score boundaries and learns one sampling weight per bin through extra bin tokens appended to the keys and values of the attention computation. The paper reports that this yields a better edge/global balance than APES in qualitative comparisons and better numbers on classification, part segmentation, and few-point sampling.","pith_inferences":["On real-world scans with non-uniform density, the kNN asymmetry signal would conflate edge-ness with density; a density-normalized variant of the sparse column count is a natural extension that could be tested on scanned-object benchmarks.","The learned bin boundaries essentially define a shape-specific importance quantile function; they could be exported as a compact descriptor of a shape class and used for rule-based sampling without neural inference.","The same sparse-attention plus bin-token design should transfer to other permutation-invariant operators, such as downsampling for point cloud registration or scene understanding, where edge preservation and coverage both matter.","The temperature τ in in-bin sampling interpolates between top-M and uniform sampling; per-bin or per-shape temperature schedules might improve the trade-off further."],"forward_implications":["Interpolation-based upsampling becomes usable: SAMBLE's uniformly spread samples let standard neighbor interpolation replace the cross-attention upsampler that APES required, without the accuracy drop APES shows under interpolation.","Few-point sampling works directly from the raw input: SAMBLE reports 90.18% overall accuracy at M=256 and 89.45% at M=32 on ModelNet40 classification without FPS pre-processing, outperforming APES that uses FPS to reduce the input first.","Sampling strategies are shape-specific and category-consistent: visualization shows similar bin histograms within a category and different strategies across categories, implying the sampler learns meaningful per-shape importance profiles.","The sampled sub-point clouds themselves carry more task signal: SAMBLE beats PointNeXt's FPS-based downsampling on intermediate downsampled sub-point clouds, suggesting the learned points, not the upsampling layer, drive the performance.","The method scales to any target sample size M through the bin-quota redistribution algorithm, so the same trained strategy can serve multiple sampling budgets."],"supporting_citations":[{"why":"APES is the immediate predecessor whose edge over-sampling and lack of shape-specific strategy SAMBLE directly targets; it is the main baseline throughout.","marker":"[46]"},{"why":"Farthest point sampling is the standard classical downsampler that SAMBLE compares against and often outperforms, especially in few-point settings.","marker":"[8]"},{"why":"S-Net, an early learning-to-sample baseline that generates new point coordinates, serves as a comparison in the few-point experiments.","marker":"[7]"},{"why":"SampleNet, a differentiable soft-projection sampler, provides a learning-based baseline that SAMBLE is compared with on few-point tasks.","marker":"[14]"},{"why":"PointNet++ supplies the hierarchical architecture and interpolation-based upsampling context against which the sampler's uniformity is evaluated.","marker":"[31]"},{"why":"The attention formulation with query, key, value, softmax, and scaling underlies the local and global attention maps from which the sparse attention map is carved.","marker":"[39]"}],"fun_headline_variants":["SAMBLE: shape-aware point cloud sampling that balances detail and uniformity","SAMBLE: per-shape sampling strategy keeps edges and global coverage","SAMBLE: sparse attention and bin learning for shape-specific sampling","SAMBLE: bin-based sampling improves few-point accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The kNN asymmetry frequency is assumed to reflect local geometry (edge-ness) rather than local point density; that holds on the uniformly sampled CAD meshes of ModelNet40 and ShapeNet-Part, but on real scans with irregular density the sampling scores could be confounded by density.","fun_headline_variants_meta":{"raw":{"variants":["SAMBLE: shape-aware point cloud sampling that balances detail and uniformity","SAMBLE: per-shape sampling strategy keeps edges and global coverage","SAMBLE: sparse attention and bin learning for shape-specific sampling","SAMBLE: bin-based sampling improves few-point accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3169,"prompt_tokens":934,"completion_tokens":2235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2164}},"tokens_in":550,"tokens_out":2235,"duration_ms":14918,"temperature":1.0,"reasoning_tokens":2164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:48:44.235560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SAMBLE on a real scan with visibly non-uniform point density and compare its top-score points against manually labeled true edges; if dense flat regions are scored as high as genuine sharp edges, the kNN-frequency signal is density, not geometry.","supporting_citations":[{"cited_title":"Attention-based point cloud edge sampling","cited_arxiv_id":null,"evidence_quote":"APES is the immediate predecessor whose edge over-sampling and lack of shape-specific strategy SAMBLE directly targets; it is the main baseline throughout."},{"cited_title":"The farthest point strategy for progres- sive image sampling.IEEE Transactions on Image Process- ing, 6(9):1305–1315, 1997","cited_arxiv_id":null,"evidence_quote":"Farthest point sampling is the standard classical downsampler that SAMBLE compares against and often outperforms, especially in few-point settings."},{"cited_title":"Learning to sam- ple","cited_arxiv_id":null,"evidence_quote":"S-Net, an early learning-to-sample baseline that generates new point coordinates, serves as a comparison in the few-point experiments."},{"cited_title":"SampleNet: Dif- ferentiable point cloud sampling","cited_arxiv_id":null,"evidence_quote":"SampleNet, a differentiable soft-projection sampler, provides a learning-based baseline that SAMBLE is compared with on few-point tasks."},{"cited_title":"PointNet++: Deep hierarchical feature learning on point sets in a metric space.Advances in Neural Information Processing Systems (NeurIPS), 30, 2017","cited_arxiv_id":null,"evidence_quote":"PointNet++ supplies the hierarchical architecture and interpolation-based upsampling context against which the sampler's uniformity is evaluated."},{"cited_title":"Attention is all you need.Advances in Neural Information Processing Systems (NeurIPS), 30, 2017","cited_arxiv_id":null,"evidence_quote":"The attention formulation with query, key, value, softmax, and scaling underlies the local and global attention maps from which the sparse attention map is carved."}],"review_version":1}