{"id":"a6b4e738-5a3c-42e1-80bc-ab3fcf28f4f3","arxiv_id":"2608.07757","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Incidence-aware manifold sampling improves resolution-averaged semantic segmentation of single-scan construction LiDAR by preserving thin, rarely observed structures.","lead":"This paper proposes a new way to downsample single-scan LiDAR point clouds before 3D segmentation, keeping more points on thin structures like ladders and pipes instead of wasteful flat surfaces. It reports consistent accuracy gains on a construction-site benchmark across two modern segmentation models, especially for rare and hard-to-see objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed ladder gains are exactly where the manifold smoothness assumption fails; without a normal-ablation control, the 'incidence-aware' mechanism is untested.","rationale":"The reader's weakest assumption identifies the same point: the method presupposes reliable normals and locally smooth surfaces, and this is exactly where the largest gains are reported. I agree that this is the most load-bearing assumption. If the incidence term is not actually responsible for the gains, then the paper's central claim—that acquisition-aware, incidence-based sampling improves segmentation—is misleading, even if the empirical numbers are reproducible. The proposed concrete test is decisive because it isolates the role of gamma while keeping the rest of the sampling pipeline intact. The reader's verdict is already CONDITIONAL, and this concern reinforces that conditionality; therefore no change to the verdict label is needed. The paper should be accepted only after such an ablation is added and reported, along with the previously noted missing hyperparameters and FPS comparison caveats.","tokens_in":14195,"tokens_out":6629,"duration_ms":72686,"concrete_test":"Run the PT experiment at resolution 0.09 m with three variants: (1) the proposed method with the normals used in the paper; (2) the same procedure with gamma set to a constant (gamma = 1), removing incidence information but retaining the radial scaling; (3) the same procedure with normals randomly permuted across points, preserving the marginal gamma distribution but destroying pointwise incidence orientation. Compare ladder IoU, NP-IoU, and mIoU against the grid baseline. If variants (2) or (3) retain most of the ladder gain (within about 2 points) and the NP-IoU improvement over grid, then the incidence term is not the active ingredient and the 'acquisition-aware' interpretation is unsupported. If the gains disappear only when gamma is removed, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that incidence-aware manifold sampling improves resolution-averaged segmentation, with the largest gains on ladders. The theoretical justification in Section 3.2 rests on the Proposition, which assumes M is locally smooth and that gamma(x) and r(x) vary smoothly over the local tangent plane, so that the Jacobian is approximately c(x) I2 and the area element transforms as dA' = w(x) dA. On thin, partially observed structures such as ladders, stairs, and pipes—precisely the classes with the largest reported gains—these assumptions are least secure. Single-scan LiDAR normals on such elements are noisy because local neighborhoods are sparse and contain edges or depth discontinuities; the paper never states how normals are computed. When gamma is inaccurate, the scaling x' = x sqrt(gamma) (r0/r)^beta is essentially a noisy perturbation rather than a faithful reparameterization of surface support. The method remains well-defined (as noted in Section 3.3), but the proposed mechanism no longer predicts improved balance. The empirical gains could instead arise from the radial term (r0/r)^beta or from a generic rearrangement of points that happens to thin out planar regions, independent of the incidence angle. Because the ladder IoU gains (+16.3 for PT, +16.9 for PNxt) are a headline result, the paper needs to demonstrate that the incidence term gamma is causally responsible, not just correlated with the selected point pattern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an incidence-aware sampling method for single-scan LiDAR point clouds. Points are mapped to a geometry-normalized space via a radial scaling that depends on the incidence term gamma(x) = |n(x)^T v(x)| and range r(x), then standard voxel-based selection is performed in that transformed space while the original Euclidean coordinates are retained for downstream segmentation. The method is evaluated on the SIP benchmark (an indoor construction-site single-scan LiDAR dataset) using Point Transformer and PointNeXt under a fixed per-fragment point budget across five sampling resolutions. The central empirical claim is that the proposed sampling improves resolution-averaged mIoU (by 4.7 points for PT and 3.1 points for PNxt) and NP-IoU, with the largest gains on ladders, and that it reduces sensitivity to sampling resolution. The paper also compares against farthest-point sampling and a lattice-preserving variant, and observes that pure manifold sampling hurts a sparse-convolution backbone (SpUNet).","tokens_in":14422,"tokens_out":4050,"duration_ms":40520,"significance":"If the empirical claims hold, the paper makes a useful case that point-allocation strategy is an active component of single-scan 3D segmentation rather than a neutral preprocessing step. The strengths include a controlled fixed-budget evaluation protocol, repeated runs with mean and standard deviation, two point-based backbones, a publicly available dataset and code, and a lightweight front-end method that requires no backbone modification. The resolution-averaged gains, particularly for non-planar classes, are non-trivial. However, the significance is tempered by the lack of an ablation isolating the incidence term from the radial term, the absence of reported values for the method's free parameters (beta and r0), and the small number of test scans underlying the headline ladder gains.","major_comments":[{"comment":"The Proposition in Section 3.2 is essentially a restatement of the mapping definition: it derives the area-element scaling w(x) = gamma(x)(r0/r)^2 from the local Jacobian approximation, but it does not establish the claimed 'more consistent correspondence between occupied cells and represented surface support.' The proof shows only that the area element transforms; the segmentation benefit is an empirical hypothesis. More importantly, the experiments do not isolate the incidence term from the radial term. The mapping in Algorithm 1 includes both gamma and (r0/r)^beta, and the radial term alone could reproduce the density redistribution that drives the gains. An ablation with gamma set to 1 (or with beta = 0) is needed to support the 'incidence-aware' attribution and to explain the ladder results in Table 6.","section":"Section 3.2 and Table 6"},{"comment":"The values of the free parameters beta and r0 are never reported in the experimental section. Algorithm 1 takes both as inputs, and Section 3.3 states that beta controls the strength of radial normalization and that h is interpreted at reference range r0, but the manuscript does not say what values were used for beta and r0 in the five-resolution experiments. Without these values, the results cannot be reproduced, and the sensitivity of the method to these parameters is unknown. Please report the chosen values and, ideally, a sensitivity study over beta.","section":"Section 3.3, Algorithm 1, and Table 3"},{"comment":"The method critically depends on surface normals, because gamma = |n^T v| is computed from unit normals. The paper never states how normals are obtained for the SIP dataset: whether they are provided by the sensor, estimated from neighborhoods, or computed with a specific estimator and radius. Since the incidence term is the core of the proposed mechanism, the normal estimation procedure (and any preprocessing such as orientation consistency) must be described. A normal-ablation study would also address the concern that the ladder gains are driven by noisy normals on thin structures.","section":"Section 3.1 and Algorithm 1"},{"comment":"The paper states that a validation set is not feasible due to rare-class coverage, but this means the hyperparameters of the proposed method (beta, r0, and any design choices) are not selected on held-out data. The test set contains only 6 scans, and the ladder class appears in only 3 test scans (Table 2). The headline resolution-averaged ladder IoU gains of +16.3 and +16.9 points in Table 6 are therefore based on a very small number of test instances. Please report per-scan ladder IoU or a bootstrap confidence interval for the resolution-averaged ladder gain, and clarify whether the reported setting corresponds to any hyperparameter search on the test split.","section":"Section 4.2 and Tables 2, 6"}],"minor_comments":[{"comment":"The notation 'r04 to r09' is used in the text but never defined; please define r04 = 0.04 m, r06 = 0.06 m, etc., at the first use.","section":"Section 5.2"},{"comment":"The table headers use 'mIOU' while the text uses 'mIoU'; please standardize the capitalization throughout.","section":"Tables 4 and 5"},{"comment":"The caption refers to 'Manifold (ours)', but the method is not explicitly named in the text; consider introducing a short name (e.g., 'incidence-aware sampling, IAS') and using it consistently in captions and tables.","section":"Figure 6"},{"comment":"The footnote states that FPS is evaluated only on sampled points while other methods are evaluated after inverse mapping to the original scan. This makes the comparison not apples-to-apples; please explain why FPS cannot be inverse-mapped and discuss how this affects the comparison.","section":"Table 7(a)"},{"comment":"Line 7 uses spherical bin indices B_r, B_theta, B_phi and a hash K, but these symbols are not defined anywhere. Please clarify the hashing/voxelization step or refer to a standard grid hashing formulation.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on the SIP benchmark [13], which appears to be a separate arXiv preprint by the same author group. The relationship between this submission and the SIP dataset paper should be clarified, particularly whether the dataset paper is under review and how the current paper's contributions are distinct. Also, the use of an unpublished benchmark may be a scope concern for a general CV venue; the authors should make the dataset and code availability explicit early in the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper ships a concrete, cheap sampling method that genuinely improves resolution-averaged segmentation on the SIP construction benchmark with two point-based backbones, and it is honest about where it fails (SpUNet). Second, the theoretical framing in Section 3.2 is a derivation by construction, not a theorem, and the paper never isolates whether the incidence term or the radial term drives the gains.\n\nWhat's actually new: the specific combination of incidence weighting and range normalization before voxel-based selection. I looked at the related work and did not find this exact recipe. The empirical protocol is a cut above the usual: five runs per cell, mean ± std in Tables 4 and 5, fixed per-fragment budget, code and data links provided. Credit where due: this is a reproducible engineering contribution.\n\nThe soft spots are real but not fatal. Most importantly, r0 and beta are never reported, and those are the two free parameters of the method. Without them, a practitioner cannot reproduce the sampling. Table 7's FPS comparison is not apples-to-apples: FPS is evaluated on sampled points only, while the others are inverse-mapped to the full scan. The authors flag that in a footnote, but the table lacks error bars and the comparison still invites the wrong reading. The Proposition is essentially a restatement of the transformation definition: they define x' = x sqrt(gamma)(r0/r)^beta and then show the area element rescales by gamma(r0/r)^2. It doesn't bear the weight of 'proposition' but it also doesn't hurt the method.\n\nThe stress-test concern about normals lands. Ladders are thin and partially observed, so single-scan normals there are noisy. The authors note the local geometric interpretation degrades near edges, but they don't provide an ablation with gamma removed or with random normals. Without that, the claim that 'incidence-aware' is the active ingredient is supported only by correlation, not causation. The radial term (r0/r)^beta could plausibly do most of the work.\n\nWho this is for: researchers working on point cloud sampling, LiDAR pre-processing, or construction-domain segmentation. It deserves a serious referee, but the review should demand the missing hyperparameters, an apples-to-apples FPS comparison, and an ablation isolating gamma. I'd lean accept after major revision, not because the theory is profound but because the empirical result is useful and honestly reported.","headline":"Useful sampling front-end for single-scan LiDAR with solid repeated-run evidence, but the mechanism is under-tested and key hyperparameters are missing.","tokens_in":14983,"tokens_out":2194,"would_cite":true,"duration_ms":20681,"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":"Downsampling individual LiDAR scans should be an active part of 3D segmentation, and resampling by incidence and range improves thin-object segmentation such as ladders.","keywords":["3D semantic segmentation","single-scan LiDAR","point cloud downsampling","incidence-aware sampling","construction site scenes","long-tailed classes","manifold voxelization","SIP benchmark"],"falsifier":"Replace the scanner-provided normals in the SIP experiments with normals estimated from local neighborhoods; if the reported ladder IoU gains of roughly 16 points disappear or reverse, the dependence on reliable normals would be demonstrated.","tokens_in":13944,"feed_emoji":"🏗️","tokens_out":9307,"duration_ms":81272,"temperature":0.7,"pith_summary":"The paper's central claim is that when individual LiDAR scans are downsampled, the sampling step itself is a source of representation bias, and that this bias can be corrected before learning by reweighting point allocation according to how each surface was seen by the sensor. It proposes an incidence-aware manifold sampling scheme that moves each point by $\\sqrt{\\gamma}(r_0/r)^\\beta$ in a normalized coordinate space, voxelizes there, and then keeps the original Euclidean coordinates of the selected points. On the SIP construction benchmark, the scheme improves resolution-averaged segmentation for two hierarchical point-based backbones, with the largest gains for ladders and other non-planar elements, and reduces sensitivity to the choice of sampling resolution. If the claim holds, point allocation should be treated as an active component of single-scan 3D segmentation rather than as a neutral preprocessing step.","feed_headline":"Sampling choice lifts thin-object 3D segmentation by 16 points","feed_subtitle":"Resampling single LiDAR scans by incidence and range improves segmentation across resolutions on the SIP construction benchmark.","key_machinery":"The load-bearing object is the incidence-aware manifold mapping $T(x)=x\\sqrt{\\gamma(x)}(r_0/r(x))^\\beta$, combined with voxel-based cell selection in the transformed space. It converts sensor-centered density bias into a coordinate-space correction: because the local area element in the transformed space satisfies $dA'=w(x)dA$ with $w=\\gamma(r_0/r)^2$, a standard voxel grid in the transformed space gives each occupied cell a more uniform share of surface support. The original coordinates are never replaced, so downstream geometric context is preserved.","core_discovery":"On the paper's own terms, the discovery is that acquisition geometry, measured by the incidence term $\\gamma(x)=|n(x)^\\top v(x)|$ and range $r(x)$, can be folded into the sampling coordinate system so that voxel cells correspond more evenly to observed surface area. The proposed mapping is $x'=x\\sqrt{\\gamma(x)}(r_0/r(x))^\\beta$; voxelization in the transformed space selects points, while the original Euclidean coordinates are fed to the network. Under the smooth-surface assumption, the Jacobian on the tangent plane is approximately $c(x)I_2$ with $c(x)=\\sqrt{\\gamma(x)}r_0/r(x)$, so the area element rescales by $w(x)=\\gamma(x)(r_0/r(x))^2$, which is the mechanism claimed to counter the over-representation of grazing and near-range points. In the SIP experiments at five resolutions, manifold sampling raises resolution-averaged mIoU from 64.0 to 68.7 for Point Transformer and from 41.6 to 44.7 for PointNeXt, with ladder IoU gains of 16.3 and 16.9 points respectively.","pith_inferences":["Beyond the paper: the same incidence-and-range reweighting could be applied to other sparse-sensor tasks such as 3D object detection or instance segmentation, where thin structures are also underrepresented.","Beyond the paper: the parameters $\\beta$ and $r_0$ are hand-set, so a data-driven or sensor-calibrated choice of these values could adapt the method to other scanner placements; the paper does not test this.","Beyond the paper: the reported confusions between stair railings, pipes, and ladders suggest that sampling alone cannot separate classes with similar local geometry, so combining incidence-aware allocation with height or contextual cues is a natural next step."],"forward_implications":["Sampling becomes an input-design decision: under a fixed per-fragment point budget, incidence-aware allocation raises resolution-averaged mIoU by 4.7 points for Point Transformer and 3.1 points for PointNeXt.","Thin and sparsely represented elements benefit disproportionately: resolution-averaged ladder IoU rises by 16.3 points for Point Transformer and 16.9 points for PointNeXt, and non-planar NP-IoU by 8.1 and 5.5 points.","The method stabilizes performance across sampling resolutions; the clearest case is the coarsest tested setting, where Point Transformer mIoU at 0.18 m improves from 53.5 with grid sampling to 65.9 with manifold sampling.","Because the front end needs only point coordinates and normals and keeps the original Euclidean coordinates of selected points, it can be inserted into existing point-based segmentation pipelines without changing the backbone.","The gain is not architecture-independent: with the sparse-convolution model SpUNet, pure manifold sampling lowers mIoU, while a mixed variant that retains grid samples roughly restores the baseline, so voxel-lattice backbones need a lattice-preserving sampling variant."],"supporting_citations":[{"why":"Provides the single-station terrestrial LiDAR scans and long-tailed class distribution that define the SIP evaluation conditions.","marker":"[13]"},{"why":"Supplies one of the two hierarchical point-based backbones whose resolution-averaged segmentation is improved by the sampling method.","marker":"[33]"},{"why":"Supplies the second backbone, showing that the gains are not specific to one feature-aggregation design.","marker":"[43]"},{"why":"Provides the reference implementation of the Point Transformer training and evaluation pipeline into which the sampling method is inserted.","marker":"[44]"},{"why":"Supplies the sparse-convolution backbone used to show that pure manifold sampling can hurt voxel-lattice models and that a lattice-preserving variant recovers the baseline.","marker":"[47]"}],"fun_headline_variants":["Resampling by incidence lifts thin-object LiDAR segmentation","Incidence-aware sampling adds 16 points to ladder segmentation","Sampling choice beats density noise in single-scan 3D segmentation","Geometry-aware sampling stabilizes LiDAR segmentation across resolutions","Incidence-aware voxel sampling boosts construction scan segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that surface normals are trustworthy and that the surface is locally smooth enough for incidence and range to vary slowly across each neighborhood; the places where the method gains the most, such as ladders, are exactly where single-scan normals are noisiest and surfaces are most fragmented.","fun_headline_variants_meta":{"raw":{"variants":["Resampling by incidence lifts thin-object LiDAR segmentation","Incidence-aware sampling adds 16 points to ladder segmentation","Sampling choice beats density noise in single-scan 3D segmentation","Geometry-aware sampling stabilizes LiDAR segmentation across resolutions","Incidence-aware voxel sampling boosts construction scan segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3625,"prompt_tokens":1006,"completion_tokens":2619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2538}},"tokens_in":622,"tokens_out":2619,"duration_ms":18524,"temperature":1.0,"reasoning_tokens":2538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:19:01.955346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the scanner-provided normals in the SIP experiments with normals estimated from local neighborhoods; if the reported ladder IoU gains of roughly 16 points disappear or reverse, the dependence on reliable normals would be demonstrated.","supporting_citations":[{"cited_title":"Automated continuous construction progress monitoring using multiple workplace real time 3D scans,","cited_arxiv_id":null,"evidence_quote":"Provides the single-station terrestrial LiDAR scans and long-tailed class distribution that define the SIP evaluation conditions."},{"cited_title":"Pointnext: Revisiting pointnet++ with improved training and scaling strategies,","cited_arxiv_id":null,"evidence_quote":"Supplies the second backbone, showing that the gains are not specific to one feature-aggregation design."},{"cited_title":"Sonata: Self-supervised learning of reliable point representations,","cited_arxiv_id":null,"evidence_quote":"Provides the reference implementation of the Point Transformer training and evaluation pipeline into which the sampling method is inserted."},{"cited_title":"Point-bert: Pre-training 3d point cloud transformers with masked point modeling,","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse-convolution backbone used to show that pure manifold sampling can hurt voxel-lattice models and that a lattice-preserving variant recovers the baseline."}],"review_version":1}