{"id":"a2c293c0-e385-480f-827f-f4d04b9a9312","arxiv_id":"2504.14017","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Raw and compressed SemanticKITTI LiDAR frame sizes can be modeled by simple distributions, and these models produce network results close to real data in ns-3.","lead":"This paper fits statistical distributions to the sizes of raw and compressed LiDAR point clouds from the SemanticKITTI dataset, and shows that these distributions reproduce network latency and throughput in ns-3 simulations. The value is an 18x faster way to simulate V2X traffic without storing or replaying large sensor datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"E2E validation shows network metrics are insensitive to frame-size distribution, so the accuracy claim is not established.","rationale":"The reader's verdict is CONDITIONAL, and the concerns I identify reinforce that conditional status rather than overturn it. The reader's weakest_assumption focuses on the constant generation interval and same-dataset evaluation; my concern is closely related but sharper: the paper's own D0/S0 result is evidence that the network-level metrics are insensitive to distribution shape, so the E2E comparison cannot validate the fitted distributions. This is load-bearing because the central claim is that statistical models are a valid alternative 'maintaining high accuracy.' If network metrics are insensitive to shape, then the accuracy claim is unsupported by the presented experiments, even though the practical speedup and mean-rate reproduction may still be useful. I would not reject the paper: the statistical fitting methodology is standard, the ns-3 module is released, and the speedup is concrete. However, the authors should either add a degenerate-model sensitivity baseline or explicitly re-scope the claim to say that aggregate network metrics depend mainly on mean frame size, not on the exact distribution. The reader's CONDITIONAL verdict is therefore appropriate; I see no reason to change it.","tokens_in":20326,"tokens_out":3825,"duration_ms":37933,"concrete_test":"Run the same ns-3 scenario (Table IV) replacing StatisticalTraffic's fitted distributions with (a) a constant burst size equal to the sample mean of each DX/SY configuration and (b) an exponential distribution with the same mean, keeping the inter-burst interval at 100 ms. Compare E2E throughput and latency against the SemanticKITTI baseline. If both degenerate models reproduce the baseline to within the same tolerance as the fitted models, then the E2E comparison cannot discriminate distribution shape, and the paper's accuracy claim must be re-scoped to aggregate traffic statistics rather than distributional fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The end-to-end validation does not establish that the fitted distributions are accurate traffic models; it only establishes that the simulated network metrics are insensitive to the distribution of frame sizes. The strongest evidence is internal to the paper: D0/S0 failed the distribution-level KS test (Table I, with no passing family), yet its E2E throughput and latency curves still nearly perfectly overlap the real-data baseline (Figs. 6-7), and its E2E KS test passes at most distances. If a statistically rejected model matches the network metrics, then those metrics are driven by aggregate quantities (mean frame size and the fixed 100 ms inter-burst interval) rather than by the shape of the size distribution. Consequently, the claim that replacing real data with statistical traffic models has a negligible impact on the network, and the conclusion that StatisticalTraffic maintains high accuracy, are not actually supported by the E2E comparison. A constant-size model with the correct mean would likely pass the same tests. The fitted distributions may still be useful, but their predictive value for network behavior is not demonstrated by the experiments as designed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes statistical distribution models for the size of LiDAR point clouds, covering one raw configuration and six HSC compression configurations derived from the SemanticKITTI dataset. The fitting procedure uses maximum likelihood estimation followed by a Kolmogorov-Smirnov test with parametric bootstrap resampling; six of the seven fitted models pass the distribution-level test, while the raw-data model D0/S0 fails. The authors implement the selected distributions in an ns-3 module called StatisticalTraffic and evaluate end-to-end throughput and latency against real SemanticKITTI traces in a 28 GHz NR V2X uplink scenario. They report near-identical network metrics, an 18x average simulation speedup, and conclude that the statistical traffic models are a valid alternative to real data while maintaining high accuracy.","tokens_in":20505,"tokens_out":4642,"duration_ms":43410,"significance":"If the claims were fully established, the paper would provide a useful practical contribution: closed-form traffic models for automotive LiDAR data that remove the need to store and process an 80 GB dataset in ns-3 simulations. The statistical machinery is standard and correctly motivated: the parametric bootstrap KS procedure addresses the fact that parameters are estimated from the data, and the implementation validation against SciPy in Table III is a good practice. The public release of the code is also a strength. However, the end-to-end validation in Section VI does not, as designed, establish that the fitted distributions are accurate traffic models; it shows that the simulated network metrics are largely insensitive to the frame-size distribution. This distinction is central to the paper's main claim, so the validation argument needs reworking before the results can be accepted at face value.","major_comments":[{"comment":"The end-to-end agreement does not support the claim that the fitted distributions accurately model LiDAR traffic. The D0/S0 model fails the distribution-level KS test in Table I, yet its E2E throughput and latency curves still overlap the real-data baseline and its E2E KS test passes at nearly all distances. This is direct evidence that the network metrics are driven by aggregate quantities, namely the mean frame size and the fixed 100 ms inter-burst interval, rather than by the shape of the frame-size distribution. To make the claim load-bearing, the paper needs a sensitivity baseline: a constant-size model with the correct mean, or a deliberately misspecified distribution, should be run through the same E2E pipeline. If such a baseline also matches, then the conclusion in Sec. VI-B2, \"replacing real data with statistical traffic models has a negligible impact on the network,\" supports network-metric robustness but not model accuracy. Without this baseline, the sentence in Sec. VI-B3 that the E2E KS results are \"another demonstration of the accuracy of the selected models\" is not justified.","section":"Sec. VI-B2 and VI-B3; Figs. 6-7; Tables VII-VIII"},{"comment":"The validation is a same-data consistency check: the distributions are fitted to SemanticKITTI in Sec. IV-B and the E2E baseline is the same SemanticKITTI traces in Sec. VI-A. There is no held-out portion of the dataset or independent data source, so the experiments do not establish generalization to other sensors, scenes, frame rates, or compression pipelines. Additionally, Sec. IV opens by assuming a constant generation interval and asserts that frame size is the most relevant traffic characteristic; temporal correlation and frame-timing jitter are not modeled. The paper should either add a held-out or cross-dataset evaluation, or explicitly restrict the claims to SemanticKITTI-like periodic LiDAR traffic. Without this, the word \"comprehensive\" in the conclusion overstates the scope of the characterization.","section":"Sec. IV and Sec. VI-A"},{"comment":"The E2E KS tests are not described with enough methodological detail. It is unclear whether the test compares paired simulation runs or aggregated distributions, how many simulation repetitions per distance are used, what significance level is adopted, and whether the parametric bootstrap recalibration from Sec. IV-A is applied. This matters because these tables are used to support the accuracy claim. Please specify the exact testing procedure, or alternatively present Tables VII and VIII as descriptive overlap indicators rather than formal goodness-of-fit results.","section":"Sec. VI-B3; Tables VII and VIII"}],"minor_comments":[{"comment":"The term \"Kolmogorov-Smirnoff\" should be corrected to \"Kolmogorov-Smirnov\" throughout.","section":"Abstract, Sec. I, Fig. 1"},{"comment":"The text says the size of LiDAR point clouds can be represented by \"Logicstic\" distributions; this should be \"Logistic.\"","section":"Sec. IV-B"},{"comment":"The text refers to \"IEEE 802.1p\" when comparing peak nominal throughput; this appears to be a typo for \"IEEE 802.11p.\"","section":"Sec. VI-B2"},{"comment":"The figures show confidence intervals but the number of simulation runs and the random seed handling are not stated; please add these details for reproducibility.","section":"Figs. 6 and 7"},{"comment":"Equation (11) uses an infinite power series for the inverse CDF of the tLocationScale distribution; please state the truncation criterion or numerical tolerance used in the implementation.","section":"Sec. V-B1"},{"comment":"The sentence \"there is an almost perfect overlap between the two sets of curves\" is stronger than the evidence supports, given that D0/S0 failed the distribution-level test; a more measured phrasing such as \"the curves are visually similar\" would be more appropriate.","section":"Sec. VI-B2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the code release is a positive feature. My main concern is not the fitting methodology, which is standard and carefully executed, but the interpretability of the end-to-end validation: as currently designed, it cannot distinguish between accurate traffic models and network metrics that are insensitive to frame-size distributions. I believe this can be fixed within the scope of the manuscript by adding a sensitivity baseline and reframing the accuracy claims, so I recommend major revision rather than rejection. I have no concerns about novelty disclosure or citation patterns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the first set of closed-form distribution fits for raw and HSC-compressed LiDAR point-cloud sizes from SemanticKITTI, plus a released ns-3 module and a concrete 18x simulation speedup. That is genuinely useful to the V2X simulation community: it lets people draw frame sizes from a parametric distribution instead of storing and replaying 80GB of traces. The bootstrap-corrected KS procedure is standard and well executed; Tables I and II are internally coherent; the implementation validation against scipy (Table III) is a nice touch. Credit is due for shipping the code.\n\nThe soft spot is the end-to-end validation, and it is a real one. The paper concludes that replacing real data with statistical traffic models has negligible impact on the network, and calls StatisticalTraffic a valid alternative maintaining high accuracy. But their own Table I shows D0/S0 fails the distribution-level KS test, and Figs. 6-7 show D0/S0 still matches the real-data throughput and latency curves almost exactly, with its E2E KS test passing at most distances. That is not evidence that the fitted distribution is accurate; it is evidence that these network metrics are insensitive to the shape of the frame-size distribution. A constant-size model with the correct mean would probably pass the same tests. The authors themselves note D0/S0 is 'sufficiently precise in terms of network metrics,' but they do not take the next step: the E2E comparison, as designed, cannot distinguish good traffic models from bad ones. That undercuts the accuracy claim.\n\nA second, smaller issue: Sec. VI-B3 says the only exception is D0/S0 for d>150m, but Table VIII shows D11/S0 failing latency at 165m and D14/S0 failing at 225 and 255m. The prose overstates the results.\n\nThe same-dataset fitting and evaluation is mild circularity. It does not invalidate the distribution fits, but it means generalization to other sensors, scenes, or variable frame timing is untested. Anyone using these models should treat them as SemanticKITTI-specific until shown otherwise.\n\nOverall: the statistical modeling and the ns-3 module are worth having. The E2E validation needs a sensitivity baseline—for example, a constant-mean model and a deliberately wrong distribution—to show the metrics can actually detect differences. Without that, the accuracy claim should be softened to 'network performance is insensitive to the size distribution.' This deserves serious peer review with major revision. I would bring it to reading group; I would probably cite the distribution fits if I work in V2X simulation.","headline":"Useful distribution fits and a released ns-3 module, but the end-to-end validation is too weak to support the 'high accuracy' claim—the metrics are insensitive to distribution shape.","tokens_in":21133,"tokens_out":2225,"would_cite":true,"duration_ms":18822,"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":"Seven fitted probability distributions for LiDAR frame sizes can replace real sensor traces in V2X network simulations, yielding nearly identical end-to-end latency and throughput while cutting run time by roughly 18 times.","keywords":["LiDAR point clouds","statistical traffic modeling","V2X communication","Kolmogorov-Smirnov test","bootstrap resampling","5G NR V2X","point cloud compression","network simulation"],"falsifier":"Run the same end-to-end simulator with frame sizes drawn from the fitted distributions on one side and frame sizes extracted from a LiDAR dataset that was not used in fitting (different sensor, scene, weather, or with jittered generation times) on the other; if the throughput or latency distributions diverge beyond the tolerances reported in the paper's KS tables, the claim that statistical models can replace real automotive data fails for that setting. A cheaper check is to apply the same bootstrap KS test to the reported distribution families against a second dataset's empirical frame-size CDF.","tokens_in":20070,"feed_emoji":"🚗","tokens_out":8131,"duration_ms":72852,"temperature":0.7,"pith_summary":"The paper aims to establish that the size of LiDAR point clouds, the dominant data load in cooperative perception for autonomous driving, can be replaced inside network simulations by a small set of fitted probability distributions without changing conclusions about latency or throughput. It fits seven distributions to frame-size data, under raw and compressed configurations produced by a semantic compression pipeline, and validates each with a bootstrap-corrected Kolmogorov–Smirnov test. It then compares end-to-end network metrics when a simulator draws frame sizes from the fitted distributions versus when it replays the real sensor frames. The curves almost perfectly overlap, the runtime drops by an average factor of about 18, and even the one model that fails the statistical goodness-of-fit test produces matching network metrics. If the claim holds, researchers can run large V2X simulation campaigns without storing, loading, and compressing multi-gigabyte sensor datasets.","feed_headline":"Statistical models make V2X simulations 18x faster","feed_subtitle":"Frame-size distributions reproduce real LiDAR latency and throughput, letting researchers skip dataset replay.","key_machinery":"The load-bearing mechanism is a parametric bootstrap-corrected Kolmogorov–Smirnov test applied to maximum-likelihood-fitted candidate distributions. Because the parameters are estimated from the same sample that is later tested, the standard KS statistic is no longer distribution-free; the paper draws one thousand bootstrap resamples from each fitted model, recomputes the KS statistic on each, and uses the 0.99 percentile of those resampled statistics as the critical value. This procedure selects the seven reported distributions, and the same fitted distributions are then encoded into a custom application-layer traffic generator so that sampled frame sizes are injected into a full-stack millimeter-wave V2X simulation alongside the real-data replay baseline.","core_discovery":"The central discovery is that a handful of parametric families—tLocationScale for raw frames and for one compressed setting, plus Normal, Gamma, Nakagami, and Logistic for the others—captures the frame-size distribution of LiDAR point clouds closely enough to be statistically interchangeable with real traces at the network level. Six of the seven fitted models pass the bootstrap Kolmogorov–Smirnov test at significance level 0.01; the raw-data model fails that test, yet in end-to-end simulation its throughput and latency still match the real-data baseline. The paper concludes that the statistical traffic model is a valid alternative to trace replay, preserving accuracy while delivering substantial simulation speedup, and that the only visible divergence appears beyond roughly 150 meters, where the radio channel is largely non-line-of-sight and statistically unstable.","pith_inferences":["The fitted distribution families are derived from a single LiDAR dataset, so a natural next test is to refit the same families on data from a different sensor, weather condition, or urban layout and check whether the families persist with only parameter changes or entirely different families are needed.","The paper assumes a constant generation interval, so extending the model to include a stochastic inter-burst process would test whether the network-level equivalence survives variable frame timing, jitter, or dropped frames.","The reported average 18x speedup arises in a workflow where real data must be read from disk and compressed during the simulation; the advantage of statistical models would shrink if traces were preprocessed and preloaded, meaning the speedup is a property of that workflow rather than an intrinsic label for the models.","Because the raw-data distribution fails the KS test yet still matches network metrics, a simpler or cheaper distribution than the heavy-tailed tLocationScale might suffice for workload generation, suggesting an experiment that measures how much distributional accuracy is actually needed for faithful network results."],"forward_implications":["Researchers can replace trace replay with the fitted distributions in V2X and cooperative-perception studies, obtaining comparable end-to-end latency and throughput while cutting simulation run time by an average factor of 18.","Raw LiDAR at 10 Hz consumes about 256 Mbps at the source, beyond the capacity of legacy V2X links, while the compressed configurations reduce the source rate to as low as about 1.3 Mbps, so the models allow fast exploration of the compression versus network-load trade-off.","The latency stays below the 100 ms autonomous-driving requirement for all configurations except the one that adds semantic-segmentation inference to moderately compressed frames, and only at distances beyond about 100 meters.","Even a model that fails a distributional goodness-of-fit test can yield accurate network-level metrics, so protocol designers can select traffic models by their end-to-end impact rather than by distributional fit alone."],"supporting_citations":[{"why":"Provides the real LiDAR point clouds whose frame-size statistics are fitted and later replayed as the baseline for network comparisons.","marker":"[11]"},{"why":"Defines the semantic compression pipeline whose seven configurations produce the compressed frame sizes being modeled.","marker":"[18]"},{"why":"Supplies the Kolmogorov-Smirnov statistic and its null-hypothesis distribution for the goodness-of-fit test.","marker":"[50]"},{"why":"Justifies the parametric bootstrap resampling needed because distribution parameters are estimated from the observed data before testing.","marker":"[51]"},{"why":"Supplies the bursty application model on which the statistical traffic generator is built.","marker":"[57]"},{"why":"Provides the millimeter-wave protocol-stack module used for the end-to-end simulation.","marker":"[58]"},{"why":"Defines the street-canyon channel model that produces the distance-dependent SINR and packet-reception statistics.","marker":"[62]"},{"why":"Sets the 100 ms latency requirement that anchors the conclusion about autonomous-driving feasibility.","marker":"[21]"}],"fun_headline_variants":["Statistical LiDAR frames: 18x faster V2X sims, same accuracy","V2X test: statistical LiDAR data replaces real traces","Parametric LiDAR models pass KS test, slash sim time","Skip trace replay: statistical LiDAR models validated","18x speedup for V2X sims with fitted LiDAR distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The models assume that LiDAR source traffic is fully characterized by frame size at a constant generation interval, and the end-to-end test reuses the same dataset the distributions were fitted to, so the claim of negligible impact has not been shown to extend to other sensors, scenes, or variable frame timing.","fun_headline_variants_meta":{"raw":{"variants":["Statistical LiDAR frames: 18x faster V2X sims, same accuracy","V2X test: statistical LiDAR data replaces real traces","Parametric LiDAR models pass KS test, slash sim time","Skip trace replay: statistical LiDAR models validated","18x speedup for V2X sims with fitted LiDAR distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2512,"prompt_tokens":947,"completion_tokens":1565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":1474}},"tokens_in":563,"tokens_out":1565,"duration_ms":11630,"temperature":1.0,"reasoning_tokens":1474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:58:49.012325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same end-to-end simulator with frame sizes drawn from the fitted distributions on one side and frame sizes extracted from a LiDAR dataset that was not used in fitting (different sensor, scene, weather, or with jittered generation times) on the other; if the throughput or latency distributions diverge beyond the tolerances reported in the paper's KS tables, the claim that statistical models can replace real automotive data fails for that setting. A cheaper check is to apply the same bootstrap KS test to the reported distribution families against a second dataset's empirical frame-size CDF.","supporting_citations":[{"cited_title":"SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences,","cited_arxiv_id":null,"evidence_quote":"Provides the real LiDAR point clouds whose frame-size statistics are fitted and later replayed as the baseline for network comparisons."},{"cited_title":"Hybrid Point Cloud Semantic Compression for Automotive Sensors: A Performance Evaluation,","cited_arxiv_id":null,"evidence_quote":"Defines the semantic compression pipeline whose seven configurations produce the compressed frame sizes being modeled."},{"cited_title":"The Kolmogorov-Smirnov test for goodness of fit,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kolmogorov-Smirnov statistic and its null-hypothesis distribution for the goodness-of-fit test."},{"cited_title":"Goodness-of-fit tests when parameters are estimated,","cited_arxiv_id":null,"evidence_quote":"Justifies the parametric bootstrap resampling needed because distribution parameters are estimated from the observed data before testing."},{"cited_title":"An Open Framework for Analyzing and Modeling XR Network Traffic,","cited_arxiv_id":null,"evidence_quote":"Supplies the bursty application model on which the statistical traffic generator is built."},{"cited_title":"End-to-end simulation of 5g mmwave networks,","cited_arxiv_id":null,"evidence_quote":"Provides the millimeter-wave protocol-stack module used for the end-to-end simulation."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"Defines the street-canyon channel model that produces the distance-dependent SINR and packet-reception statistics."},{"cited_title":"Service requirements for enhanced V2X scenarios,","cited_arxiv_id":null,"evidence_quote":"Sets the 100 ms latency requirement that anchors the conclusion about autonomous-driving feasibility."}],"review_version":1}