{"id":"97051e09-65e5-4482-94cf-68a1c0e93535","arxiv_id":"2505.24634","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Non-uniform radial voxelization with arithmetic-progression intervals improves LiDAR semantic segmentation accuracy and efficiency on SemanticKITTI and nuScenes.","lead":"A LiDAR segmentation method that partitions space into non-uniform, distance-growing cells reports higher accuracy with four times less computation and faster training than uniform cell methods. The core efficiency-accuracy trade-off is plausible, but one headline 'state-of-the-art' claim is contradicted by the paper's own benchmark table.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 'state-of-the-art on nuScenes' is false: Table II shows SFPNet at 80.1 mIoU vs NUC-Net's 78.9 on validation, so the central SOTA claim is internally contradicted.","rationale":"I agree with the reader's overall CONDITIONAL verdict, but I identify a different load-bearing concern than the reader's weakest_assumption. The reader focused on the fairness of efficiency comparisons, which is a legitimate concern about hardware and measurement conditions. I see that as an important but somewhat speculative issue. The more decisive concern is the false SOTA claim on nuScenes validation, which is directly contradicted by the paper's own Table II. This is not an interpretation question: SFPNet is listed at 80.1 mIoU, NUC-Net at 78.9 mIoU, both LiDAR-only single-scan methods. The abstract and Section IV-C claim state-of-the-art performance on nuScenes without qualification, so the central claim fails as written. The non-uniform partition mechanism itself appears sound: Table V isolates the partition change across two backbones and three resolutions, showing consistent gains, and Equation (4) correctly implements nested coarsening of the arithmetic-progression intervals. Thus the technical contribution should survive, but the paper must correct the SOTA claim, disclose validation-set hyperparameter selection more prominently, and add error bars or multiple seeds. These corrections would still leave the paper as a conditional accept rather than an unconditional one. The efficiency comparison concern from the reader remains unresolved and should be addressed by re-measuring baselines on matched hardware, but the SOTA contradiction is the single most load-bearing issue because it is already proven by the manuscript's own data.","tokens_in":24579,"tokens_out":14441,"duration_ms":162591,"concrete_test":"Verify the SFPNet nuScenes validation mIoU of 80.1 by consulting the original SFPNet paper (ECCV 2024) or the official nuScenes benchmark results, and independently run the released NUC-Net code on the nuScenes validation split to confirm the 78.9 mIoU. If SFPNet's 80.1 is correctly transcribed and NUC-Net's 78.9 reproduces, the abstract's SOTA claim is definitively false and must be amended.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, as stated in the abstract and repeated in Section IV-C, is that NUC-Net achieves state-of-the-art performance on SemanticKITTI and nuScenes. On nuScenes validation, Table II lists SFPNet at 80.1 mIoU and NUC-Net at 78.9 mIoU; on nuScenes test, Table III shows a tie at 80.2 mIoU. SFPNet is a LiDAR-only, single-scan method, so it is a direct comparison rather than one requiring RGB input or temporal context. The paper even acknowledges in the table that the baseline and NUC-Net use only LiDAR input. Therefore, the unqualified statement 'achieves state-of-the-art performance on SemanticKITTI and nuScenes datasets' is not supported by the paper's own numbers. This is not a matter of hardware or measurement protocol, which the reader's weakest_assumption flagged; it is an internal factual inconsistency. The core non-uniform partition mechanism may still be valuable, and the matched-resolution ablations in Table V are credible, but the headline SOTA claim is a central part of the contribution and must be corrected before the paper can be accepted as is. If the claim were revised to 'comparable to state-of-the-art' or 'state-of-the-art on SemanticKITTI and tied on nuScenes test', the technical content would remain, but as written the central claim is false on nuScenes validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NUC-Net, a LiDAR semantic segmentation method that replaces the uniform radial partition of cylindrical voxel grids with a non-uniform Arithmetic Progression of Interval (API) partition, together with a non-uniform multi-scale aggregation module. The authors report state-of-the-art results on SemanticKITTI and nuScenes, with 4x faster training, 2x lower GPU memory, and 3x inference speedup relative to a uniform counterpart, and they provide a qualitative theoretical analysis based on encoding error and receptive-field arguments. The paper also includes generalization experiments to multi-scan segmentation, panoptic segmentation, scribble-annotated data, and corrupted-data robustness.","tokens_in":24711,"tokens_out":9998,"duration_ms":110544,"significance":"If the claims hold, the contribution is practically valuable: a partition change that keeps accuracy while allowing a 4x input-resolution reduction would make voxel-based LiDAR segmentation considerably cheaper. The strongest evidence is Table V, which shows clean, controlled gains from replacing uniform with non-uniform partition at identical voxel resolution for both PolarNet (+6.1 mIoU at [120,360,32]) and Cylinder3D (+4.7 mIoU at [120,360,32]). The breadth of additional experiments (multi-scan, panoptic, scribble, Robo3D) is a genuine strength, and the code is promised to be public. The theoretical analysis in Section IX is qualitative rather than formal, but the core empirical mechanism is supported by the controlled ablation.","major_comments":[{"comment":"The state-of-the-art claim on nuScenes is internally contradicted by the paper's own numbers. On the nuScenes validation set, Table II reports SFPNet at 80.1 mIoU and NUC-Net at 78.9 mIoU; on the test set, Table III shows NUC-Net tied with SFPNet at 80.2 mIoU. SFPNet is a LiDAR-only, single-scan method and is therefore a direct comparison. The abstract and Section IV-C claim \"state-of-the-art performance on SemanticKITTI and nuScenes datasets,\" which is not supported by these tables. The authors should revise the headline claim to \"comparable\" or \"competitive\" on nuScenes, or explicitly justify why SFPNet should be excluded from the comparison.","section":"Section IV-C, Tables II and III"},{"comment":"The efficiency gains attributed to the non-uniform partition are not measured at matched resolution. Table I shows the uniform \"Baseline\" at the same input resolution (120x360x32) has 52 ms inference speed, while \"Ours\" has 56 ms, so the partition change alone does not produce a 3x inference speedup. The 3x/4x/2x figures in Table XIV compare NUC-Net at 120x360x32 against Cylinder3D at 480x360x32, conflating the effect of the partition with the effect of reducing input resolution. Since Table V reports only accuracy at matched resolution, the \"general component\" claim for efficiency is unsupported. Please report inference speed, GPU memory, and training time under matched resolution for the uniform and non-uniform versions, or qualify the claims to state that the efficiency gains come from the combination of non-uniform partition and reduced input resolution.","section":"Abstract; Section IV-E, Tables I and XIV"},{"comment":"The uniform baseline jumps from 63.2 mIoU on validation (Table IV) to 69.5 mIoU on test (Table I) without explanation. Because this baseline is the control for the headline SemanticKITTI comparison, the source of the discrepancy (e.g., test-time ensemble, instance augmentation, validation/test distribution) should be stated explicitly. Without this, the reader cannot reconcile the ablation numbers with the main results or assess how much of the final gain is attributable to the proposed components.","section":"Section IV-D, Table IV vs Table I"}],"minor_comments":[{"comment":"There are several typos and inconsistencies: \"wihch\" (Section I), \"V oxel\" and \"voxel-vise\" in the Figure 2 caption, \"Arthmetric\" and \"tollerance\" in Section IV-D, \"furhter\" in Section VII, and \"conprehensive\" in Section VIII. Please proofread the manuscript.","section":"Throughout"},{"comment":"The footnote marking asterisked speeds is insufficient for comparing efficiency claims. Cylinder3D's 171 ms and RPVNet's 168 ms are not marked as RTX 3090 measurements, and RangeFormer's 37 ms is explicitly from an A100. Please report the hardware and measurement protocol for every speed entry, not only the asterisked ones.","section":"Table I"},{"comment":"The row labeled \"Salsanext [3]\" cites reference [3] (PolarNet), which appears to be a citation error; the correct reference for SalsaNext should be used.","section":"Table II"},{"comment":"The voxel counts in Table XIII are non-integers (e.g., 21015.1), which suggests these are frame-averaged values; please state this explicitly. Also, \"1.72 lower\" should be \"1.72x lower\" (or \"by a factor of 1.72\").","section":"Table XIII and surrounding text"},{"comment":"The text mentions Cylinder3D with input voxel resolution [120,360,180], but Table XII lists Cylinder3D[120,360,32]. Please align the text and table.","section":"Section VIII, Table XII"},{"comment":"The theoretical analysis is qualitative (encoding error, receptive field, sparse voxel communication). The abstract's phrase \"theoretical analysis\" overstates the level of formality; consider describing it as an empirical and qualitative analysis.","section":"Section IX"}],"recommendation":"major_revision","confidential_remarks":"The core mechanism is supported by the matched-resolution ablations in Table V, so I do not recommend rejection. However, the paper's headline claims are not supportable as written: the nuScenes SOTA claim is internally contradicted by Tables II and III, and the efficiency claims compare unequal resolutions. The authors should be asked to re-benchmark at matched resolution and to revise the abstract and contribution statements accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical methods paper with a clean core idea and an over-claimed headline. The controlled ablations in Table V show that swapping uniform for arithmetic-progression radial bins at matched resolution gives +4.7/+6.1 mIoU on Cylinder3D/PolarNet, and Eq. (4) checks out as a nested coarsening formula. That is real evidence the mechanism works. I would not dismiss this.\n\nWhat is new: the API partition along the radial axis and the scale-consistent non-uniform aggregation. The authors correctly distinguish their work from OctNet and SphereFormer, and the citation pattern looks fair. They also ship a code link, though I did not verify it runs. The theoretical section in Section IX is descriptive rather than predictive; treat it as intuition, not proof.\n\nThe problems are in the framing, not the mechanism. The abstract says state-of-the-art on nuScenes, but Table II has SFPNet at 80.1 versus NUC-Net's 78.9 on validation, and Table III is a tie on test. That is an internal factual contradiction. It must be fixed to \"comparable\" or \"state-of-the-art on SemanticKITTI and tied/competitive on nuScenes.\" Second, the efficiency claims mix measurement conditions: Cylinder3D's 171 ms is not flagged as author-measured, and RangeFormer's 37 ms is an A100 number. The 3x/4x/2x deltas are likely inflated as stated. Third, the two API hyperparameters are selected on the validation set (Section IV-D), so the reported validation gain is partly tuned. The authors disclose this, but it should be stated more clearly, and error bars or multiple seeds would help. Minor: the uniform baseline jumps from 63.2 mIoU on validation to 69.5 on test with no explanation; that is a test-versus-validation comparison and needs a clarifying sentence.\n\nWho this is for: anyone working on voxel-based LiDAR segmentation or efficient point cloud representations. It deserves a serious referee; the mechanism is worth engaging with despite the over-claims. I would ask for a revision that corrects the SOTA claims, labels hardware/protocol for every timing number, and reports variance.","headline":"Solid non-uniform partition mechanism with clean controlled ablations, undermined by an over-claimed SOTA statement and mixed-hardware efficiency comparisons; needs revision, not rejection.","tokens_in":25451,"tokens_out":1743,"would_cite":true,"duration_ms":20309,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Non-uniform radial voxels—spaced in an arithmetic progression—are claimed to make voxel-based LiDAR semantic segmentation both more accurate and several times cheaper in compute, memory, and training time than uniform cylindrical grids.","keywords":["LiDAR semantic segmentation","non-uniform partition","cylindrical coordinates","arithmetic progression of interval","sparse convolution","SemanticKITTI","nuScenes","voxel representation"],"falsifier":"Run the identical 3D sparse-convolution network on SemanticKITTI with only the partition changed—uniform radial intervals vs API—at exactly the same voxel resolution, batch size, GPU, and inference script, and measure wall-clock latency, peak GPU memory, and validation mIoU. If the API partition does not reproduce a substantial mIoU gain at matched resolution (the paper's Table V reports +4.7 for Cylinder3D at $[120,360,32]$) and does not reduce non-empty voxels or active sites in deeper layers as claimed, the central efficiency-accuracy claim fails.","tokens_in":24168,"feed_emoji":"🚗","tokens_out":6291,"duration_ms":73454,"temperature":0.7,"pith_summary":"This paper argues that the standard way of dividing 3D space for LiDAR segmentation—uniform cylindrical voxels—is both wasteful and representational: equal radial intervals put too many voxels where points are dense and too few where they are sparse. The paper proposes NUC-Net, which spaces radial intervals in an arithmetic progression, so voxels grow slowly near the sensor and rapidly far away. At a quarter of the radial resolution of the standard Cylinder3D representation, the network reports higher semantic segmentation accuracy (73.6 vs 67.8 mIoU on SemanticKITTI test) with roughly three times faster inference, about four times less training time, and about half the GPU memory. If these comparisons hold under matched measurement, a simple change to how voxels are laid out, rather than a new network family, is a general efficiency lever for voxel-based LiDAR perception. The paper also gives a mechanism for the gain: smaller near-field voxels reduce label-mixing encoding errors, while larger far-field voxels widen the receptive field of fixed-size convolution kernels.","feed_headline":"One partition change makes LiDAR segmentation faster and more accurate","feed_subtitle":"Arithmetic-progression radial intervals give 73.6 mIoU on SemanticKITTI at a quarter the voxel resolution, with 3x faster inference.","key_machinery":"The load-bearing object is the Arithmetic Progression of Interval (API) radial partition: the $i$-th radial interval is $a_i = a_0 + i\\,d$, chosen so voxel volume grows approximately as $b h d^2 i^3/2$ rather than linearly in the radial index. It controls how many voxels are allocated to dense near-field vs sparse far-field regions, determining both encoding error and receptive field under a fixed 3D sparse-convolution backbone. A second component, non-uniform multi-scale aggregation, builds multi-scale voxel features whose radial intervals are aligned so that scale stacking does not create mismatched cell boundaries. The argument is carried by the volume-growth identity $V_{i,j,k}\\propto i^3$ and the empirical point-density profile of LiDAR scans.","core_discovery":"The central claim is that replacing the uniform radial partition of cylindrical-coordinate voxelization with an Arithmetic Progression of Interval (API) partition—where the $i$-th radial interval is $a_i = a_0 + i d$—makes the voxel volume grow like $i^3$ instead of $i$, matching the way LiDAR point density falls off with distance. This yields a more balanced points-per-cell distribution, lower encoding error near the sensor, and a larger effective receptive field in distant regions without enlarging convolution kernels or raising resolution. The paper reports that this one change, applied as a general component, raises Cylinder3D from 61.8 to 66.5 mIoU and PolarNet from 52.6 to 58.7 mIoU at matched resolution on SemanticKITTI validation, and that NUC-Net reaches 73.6 mIoU on the SemanticKITTI test set at $120\\times360\\times32$ resolution with 19.9 GMACs and 56 ms inference. The same representation is shown to transfer to multi-scan semantic, panoptic, scribble-supervised, and corruption-robustness settings. The paper frames the contribution as a representation-level improvement that is complementary to efficient range-view, distillation, and lightweight-architecture methods.","pith_inferences":["The volume-growth law $V\\propto i^3$ suggests an explicit design rule the paper does not pursue: the radial interval schedule should track the inverse of the point-density profile of the sensor, so other sensors with different falloff rates could need different exponents or progressions, a testable per-sensor calibration.","The same API idea could be applied to the range (column) axis of spherical range-view representations, potentially transferring the speed-accuracy gain to 2D-convolution methods that currently dominate real-time deployments.","The paper's encoding-error argument predicts that non-uniform partition should help most in mixed-class boundary voxels near the sensor; a fine-grained evaluation of boundary IoU by distance bin would make that mechanism directly visible.","An implicit economic consequence is that if training memory indeed drops by roughly half, larger batch sizes or longer contexts such as multi-frame input become affordable without changing the model, extending the method's value beyond single-scan segmentation."],"forward_implications":["If the central claim is correct, voxel-based LiDAR segmentation does not need higher input resolution to gain accuracy; the same accuracy can be reached at one-quarter the radial resolution.","Uniform cylindrical networks such as Cylinder3D and PolarNet can be upgraded to non-uniform partition without changing their architecture, gaining accuracy and speed simultaneously.","The reported 19.9 GMACs, 56 ms inference, and 2.5 GPU-days training profile would make state-of-the-art LiDAR segmentation practical on a single consumer GPU, lowering the barrier for real-time autonomous driving systems.","The non-uniform partition transfers to multi-scan, panoptic, and scribble-annotated settings, so the benefit is not specific to one task.","Because the method is representation-level, it can be combined with range-view, distillation, and lightweight-architecture approaches that target efficiency from other directions."],"supporting_citations":[{"why":"Supplies the uniform cylindrical partition baseline, the asymmetrical 3D sparse-convolution network backbone, and the mIoU numbers that the non-uniform partition improves upon.","marker":"[1]"},{"why":"Supplies the uniform polar/cylindrical grid representation and the PolarNet baseline that gains up to 6.1 mIoU from the non-uniform partition.","marker":"[3]"},{"why":"Provides the SemanticKITTI dataset and evaluation protocol on which the main accuracy and efficiency claims are measured.","marker":"[34]"},{"why":"Provides the nuScenes dataset and 16-class protocol used to show generalization beyond SemanticKITTI.","marker":"[35]"},{"why":"Supplies the point-voxel module and the sparse-convolution MACs estimation procedure used in the efficiency analysis.","marker":"[10]"},{"why":"Supplies the SphereFormer comparison point and the exponential-splitting position encoding that the paper distinguishes from a speed-improving non-uniform partition.","marker":"[13]"},{"why":"Supplies the 2DPASS baseline and the test-time ensemble practice the paper follows during inference.","marker":"[11]"},{"why":"Supplies the PVKD distillation baseline whose training time, memory, and accuracy the paper contrasts with NUC-Net's.","marker":"[16]"}],"fun_headline_variants":["Non-uniform radial bins make LiDAR segmentation 3x faster","Arithmetic spacing of voxels triples LiDAR inference speed","Adaptive voxel spacing boosts LiDAR segmentation speed","Arithmetic bin intervals give LiDAR segmentation 3x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported speed, memory, and training-time gains assume the comparison numbers for the uniform baselines were measured under the same hardware and inference protocol as the paper's own runs; if Cylinder3D's 171 ms, PVKD's 76 ms, and other baseline timings come from different GPUs or settings, the claimed 3x, 4x, and 2x improvements shrink accordingly.","fun_headline_variants_meta":{"raw":{"variants":["Non-uniform radial bins make LiDAR segmentation 3x faster","Arithmetic spacing of voxels triples LiDAR inference speed","Adaptive voxel spacing boosts LiDAR segmentation speed","Arithmetic bin intervals give LiDAR segmentation 3x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000641,"raw_usage":{"total_tokens":3037,"prompt_tokens":1119,"completion_tokens":1918,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":1848}},"tokens_in":735,"tokens_out":1918,"duration_ms":19837,"temperature":1.0,"reasoning_tokens":1848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:19:55.972030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical 3D sparse-convolution network on SemanticKITTI with only the partition changed—uniform radial intervals vs API—at exactly the same voxel resolution, batch size, GPU, and inference script, and measure wall-clock latency, peak GPU memory, and validation mIoU. If the API partition does not reproduce a substantial mIoU gain at matched resolution (the paper's Table V reports +4.7 for Cylinder3D at $[120,360,32]$) and does not reduce non-empty voxels or active sites in deeper layers as claimed, the central efficiency-accuracy claim fails.","supporting_citations":[{"cited_title":"Cylindrical and asymmetrical 3d convolution networks for lidar segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the uniform cylindrical partition baseline, the asymmetrical 3D sparse-convolution network backbone, and the mIoU numbers that the non-uniform partition improves upon."},{"cited_title":"Polarnet: An improved grid representation for online lidar point clouds semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the uniform polar/cylindrical grid representation and the PolarNet baseline that gains up to 6.1 mIoU from the non-uniform partition."},{"cited_title":"Semantickitti: A dataset for semantic scene understanding of lidar sequences,","cited_arxiv_id":null,"evidence_quote":"Provides the SemanticKITTI dataset and evaluation protocol on which the main accuracy and efficiency claims are measured."},{"cited_title":"nuscenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes dataset and 16-class protocol used to show generalization beyond SemanticKITTI."},{"cited_title":"Search- ing efficient 3d architectures with sparse point-voxel convolution,","cited_arxiv_id":null,"evidence_quote":"Supplies the point-voxel module and the sparse-convolution MACs estimation procedure used in the efficiency analysis."},{"cited_title":"Spherical transformer for lidar-based 3d recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the SphereFormer comparison point and the exponential-splitting position encoding that the paper distinguishes from a speed-improving non-uniform partition."},{"cited_title":"2dpass: 2d priors assisted semantic segmentation on lidar point clouds,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2DPASS baseline and the test-time ensemble practice the paper follows during inference."},{"cited_title":"Point-to-voxel knowledge distillation for lidar semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the PVKD distillation baseline whose training time, memory, and accuracy the paper contrasts with NUC-Net's."}],"review_version":1}