{"id":"3193622d-f70b-4ffe-a01f-6e3bad433ab4","arxiv_id":"2507.06976","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DenoiseCP-Net jointly denoises and detects in a single LiDAR network, cutting collective-perception bandwidth by up to 23.6% in simulated adverse weather while keeping detection accuracy.","lead":"An autonomous-driving team built a network that cleans weather noise from LiDAR data while also detecting objects, so vehicles can share smaller, cleaner data packets. In simulated rain, snow, and fog, the system removes most noise voxels (over 97% accuracy) and cuts transmission bandwidth by up to 23.6% on average, with no loss in detection accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Denoising accuracy and bandwidth savings are measured only on the authors' synthetic weather noise, with labels from the same simulation; if that noise signature differs from real rain/snow/fog, the 99% F1 and 23.6% bandwidth reduction will not transfer to deployment.","rationale":"The concern is external validity, not internal inconsistency. The paper's experiments are internally plausible: the architecture is clearly described, the reported metrics are coherent, and the lack of a two-stage baseline weakens the latency claim but does not by itself falsify it. I credit the fog model's calibration to real fog-chamber data and the use of physically motivated drop-size distributions as partial support. However, the strongest claim is about performance 'in adverse weather' generally, not only under the authors' simulator. Because the same simulation produces both the phenomenon and the supervision, the near-perfect F1 and bandwidth savings could be inflated by simulator-specific signatures, such as the particular spatial distribution of noise voxels that the model can learn to recognize. This is exactly the reader's weakest_assumption, so I agree with that identification. The proposed real-data or held-out-parameter test would settle whether the concern actually lands. Conditional acceptance remains appropriate: the contribution is plausible and useful, but the real-world transfer is unverified, and the paper itself acknowledges that real-world evaluation is future work. No verdict change from the reader is needed; the conditional framing already captures this uncertainty.","tokens_in":11973,"tokens_out":6734,"duration_ms":79370,"concrete_test":"Take the trained DenoiseCP-Net (no fine-tuning) and run its denoising decoder on real adverse-weather LiDAR sequences with independent voxel-level labels—e.g., manually annotate noise voxels in a sample of the Canadian Adverse Driving Conditions (CADC) snow/rain data, or use the real fog-chamber recordings from Heinzler et al. [31]. Compute per-class accuracy and F1 for the noise/no-noise classes and compare to Table II. If the noise-class F1 falls from ~99% to below ~90%, or the no-noise F1 drops similarly, the simulated noise is not representative, and the claimed bandwidth/latency reductions are likely overestimates for real deployment. A cheaper analytical cross-check is to retrain on low-intensity simulation parameters only and test on the high-intensity parameters from Table V; a large F1 drop would confirm simulator-signature overfitting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—near-perfect denoising, up to 23.6% bandwidth reduction, and unchanged detection AP—is established empirically only on OPV2V augmented by the authors' own weather simulations [5]. The voxel-level denoising ground truth is generated by the same simulation: Section VI-A states that 'false points caused from the rain, snow, or fog simulation are then labeled noise.' This creates a closed evaluation loop: the denoising decoder can exploit spatial statistics peculiar to the simulator (e.g., isolated voxels in free space created by ray-tracing/hit-ratio thresholds) rather than learned physical properties of weather noise. The fog model has some independent support because it was fitted to real fog-chamber recordings (Sec. IV-C), and physically based drop-size distributions are used for rain and snow, but no quantitative comparison to real rain/snow LiDAR noise is provided. Since every downstream quantity—denoising F1, bandwidth savings, and the claim that removed noise carries no detection information—is conditioned on these simulated noise voxels being representative of real weather, the external-validity gap is the single most load-bearing weak point. The paper itself acknowledges in the Conclusion that real-world collective-perception weather data are not yet available, so the appropriate reading is that the claims hold in the simulation domain only.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DenoiseCP-Net, a multi-task sparse-voxel architecture for LiDAR-based collective perception that jointly performs voxel-level noise filtering and 3D object detection with a shared sparse-convolution backbone. The authors extend the OPV2V dataset by simulating rain, snow, light fog, and dense fog using their previously developed weather simulation models, and evaluate DenoiseCP-Net under three cooperative-sharing schemes: noisy, denoised, and mixed. Reported results include near-perfect per-class denoising accuracy and F1 scores (Table II), 3D object detection AP that is nearly unchanged when noisy cooperative voxel grids are replaced by denoised ones (Tables III–IV), bandwidth reductions of up to 23.6% (and up to 62.3% under high-intensity weather, Table V), and modest inference-latency reductions. The paper concludes that removing weather-induced noise before communication reduces bandwidth and computation without sacrificing detection performance.","tokens_in":12253,"tokens_out":9068,"duration_ms":100002,"significance":"The engineering idea of coupling voxel-level denoising with detection in one sparse-convolution backbone is sensible and potentially useful: it avoids a separate denoising stage, and the reported AP preservation under denoised transmission is internally consistent within the simulated domain. The paper also provides clearly described, physically motivated weather simulation models (drop-size and snowflake-size distributions, fog chamber fitting) and reports per-class metrics, which is more informative than a single aggregate score. The main limitation is external validity: all quantitative claims are produced on the authors' own simulated weather, with denoising ground truth generated by the same simulation, so the headline numbers should be read as simulation-domain results unless additional real-data evidence is supplied. If the claims are appropriately scoped, this would be a solid contribution to the collective-perception literature, but the current manuscript overstates the strength of the evidence.","major_comments":[{"comment":"","section":"VI-A (with IV-A–C)"},{"comment":"","section":"V-A / VII-C"},{"comment":"","section":"Table III"}],"minor_comments":[{"comment":"","section":"VI-C / Table II"},{"comment":"","section":"VI-B / VII-C"},{"comment":"","section":"VII-A"},{"comment":"","section":"VII-B"},{"comment":"","section":"II-B / Contributions"}],"recommendation":"major_revision","confidential_remarks":"The central technical contribution is sound within the simulated domain, but the external-validity gap is load-bearing and must be addressed before publication. The authors may be able to fix this by re-scoping the claims and adding a real-data sanity check or at least a simulation-parameter robustness study. The unsupported two-stage latency claim also needs attention. If these issues are addressed, the paper would be a reasonable contribution to the intelligent-vehicles collective-perception literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DenoiseCP-Net is a sensible engineering contribution: it combines voxel-level denoising with 3D detection in a single sparse-convolution backbone for collective perception. The architecture is clean—a shared ego backbone feeds both a U-Net denoiser and a fusion backbone that scatters ego features into the cooperative stream. On the authors' weather-augmented OPV2V, the numbers are internally consistent: denoising preserves AP, bandwidth drops by up to 23.6%, and latency improves a few percent. That is genuinely useful for the subfield.\n\nThe main thing to know is that the evaluation is entirely on the authors' own simulated weather, and the denoising ground truth comes from the same simulation that adds the noise. So the near-perfect F1 scores and the bandwidth reduction are guaranteed to hold only inside this closed loop. The fog model was fitted to real fog-chamber data, which gives it some independent grounding; rain and snow are physically based but not validated against real LiDAR noise. The authors acknowledge there is no real-world CP weather dataset, so the honest reading is: the method works well on their simulator, and we don't know how it transfers.\n\nThere are a couple of smaller soft spots. The mixed noisy/denoised results are weird in snow—AP drops from 61.29 to 57.59, well below both the noisy and denoised conditions—and the paper doesn't explain it. No error bars are reported. The 'first study' claim is a bit overstated given Jiang et al. on rain and V2X-R on fog/snow with radar, though the LiDAR-only voxel-grid framing is new.\n\nOverall, the architecture is worth taking seriously. The central limitation is external validity, not internal logic. A careful referee should ask for either a real-weather test or a much clearer statement that the claims are simulation-only, plus a check on the mixed-snow result. I'd send it to review.","headline":"A clean multi-task backbone for collective perception that delivers real bandwidth savings in simulation, but the closed-loop weather evaluation makes external validity the load-bearing question.","tokens_in":12810,"tokens_out":2259,"would_cite":true,"duration_ms":22049,"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":"DenoiseCP-Net filters simulated rain, snow, and fog from LiDAR voxel grids before sharing, keeping detection accuracy unchanged while cutting bandwidth by up to 23.6%.","keywords":["collective perception","LiDAR","adverse weather","3D object detection","point cloud denoising","sparse voxel grids","multi-task learning","bandwidth reduction"],"falsifier":"Train DenoiseCP-Net on the simulated weather-augmented OPV2V data and evaluate it on real rainy, snowy, or foggy LiDAR sequences with physically validated noise labels; if per-voxel noise F1 drops substantially below the simulated near-99% levels, or if detection AP with denoised sharing falls below AP with noisy sharing, the central claim fails. A simpler check is to compute the transmitted payload size for real adverse-weather point clouds before and after denoising; if real noise does not inflate the voxel count, the reported 23.6% bandwidth reduction will not appear.","tokens_in":11780,"feed_emoji":"🌧️","tokens_out":7286,"duration_ms":73542,"temperature":0.7,"pith_summary":"This paper tries to establish that adverse-weather noise can be stripped out of LiDAR-based collective perception without paying a detection penalty, and that doing so before transmission makes the whole system cheaper. The authors propose DenoiseCP-Net, a single sparse-convolution network that labels each occupied voxel of the ego vehicle's LiDAR grid as noise or structure and, at the same time, detects 3D objects from fused local and cooperative voxel grids. To study the problem they extend OPV2V with simulated rain, snow, and fog and show near-perfect voxel denoising (F1 above roughly 96% in dense fog and above 99% in other conditions), detection accuracy essentially unchanged compared with sharing noisy voxels, an average bandwidth reduction up to 23.6%, and reduced inference latency in cooperative vehicles. The practical payoff would be collective perception that stays accurate in bad weather and uses less communication and compute; the main condition is that the simulated weather noise resembles real sensor noise.","feed_headline":"Denoising before sharing cuts vehicle-to-vehicle bandwidth by 23.6%","feed_subtitle":"A shared LiDAR network filters rain, snow, and fog noise from voxel grids without sacrificing object detection.","key_machinery":"The central object is the shared sparse 3D convolution backbone with a U-Net-style denoising decoder and a parallel collective fusion backbone connected by a scatter operation. The shared backbone processes the ego vehicle's noisy voxel grid; its bottleneck features feed both the denoising decoder, which labels every occupied voxel noise or no-noise, and, through element-wise maximum scatter fusion, the collective fusion backbone, which ingests cooperative voxel grids from other vehicles. After training, voxels predicted as noise are deleted before the ego grid is shared, so the information transmitted is a denoised sparse voxel grid with coordinates only. This joint design is what avoids the redundant computation of a separate denoiser-plus-detector pipeline and what makes the bandwidth and latency reductions possible.","core_discovery":"On its own terms, the paper's discovery is that voxel-level weather-noise classification and 3D object detection can be done by one shared sparse-convolution backbone, and that the resulting denoised sparse voxel grid is exactly as useful to a cooperative detector as the noisy grid. On the weather-augmented OPV2V test split, DenoiseCP-Net reaches per-voxel noise accuracy of 99.73% (snow), 99.77% (rain), 99.98% (light fog), and 99.87% (dense fog), with F1 scores near the same levels. Detection AP with denoised cooperative data is essentially identical to AP with noisy data in every weather condition, for example 61.29 vs 61.29 in snow and 80.86 vs 80.82 in rain, and mixed fleets of half-denoised, half-noisy vehicles stay within a point or two. Because noise voxels are removed before transmission, the average bandwidth needed to share a vehicle's grid at 10 Hz falls by up to 23.6% (snow), and under high-intensity fog and snow it falls by 52.3% and 62.3%. The inference latency of the whole system also drops, most strongly in snow, by 6.2%.","pith_inferences":["A reader can infer that weather-adaptive sharing is possible: vehicles could decide to denoise and transmit only when the local voxel grid contains a large fraction of predicted noise, using the denoising confidence as a control signal.","The scatter-max fusion assumes all agents use the same backbone; a heterogeneous fleet of vehicles with different feature extractors would need a feature-space alignment mechanism before the reported gains transfer.","The same joint denoise-and-detect pattern should carry over to other adverse-weather perception modalities, such as camera BEV or radar voxel grids, where noise statistics differ and the bandwidth-latency trade-off may shift.","One testable extension is to run the trained model on real rainy or snowy LiDAR sequences and track per-voxel F1 and transmitted payload size; those numbers would show how much of the simulated gain survives real noise."],"forward_implications":["A cooperative vehicle can clean its own LiDAR data and detect objects in one forward pass, eliminating a separate denoising network and its redundant computation.","Weather noise carries no information needed by the downstream detector, so filtering it before sharing does not hurt AP in rain, snow, light fog, or dense fog.","The bandwidth cost of collective perception in adverse weather can be cut by up to 23.6% on average and by over half in extreme fog and snow, while inference latency also falls.","Mixed deployments, in which only some vehicles denoise before sharing, retain nearly the same detection accuracy as fully noisy or fully denoised fleets.","The same joint architecture can be applied to other dense voxel tasks such as semantic segmentation, since denoising is structurally a per-voxel segmentation problem."],"supporting_citations":[{"why":"Supplies the rain, snow, and fog simulation models that generate the weather-augmented point clouds and the voxel-level noise ground-truth labels.","marker":"[5]"},{"why":"Establishes sparse voxel grids with coordinate-only sharing as the compact representation that DenoiseCP-Net transmits and denoises.","marker":"[4]"},{"why":"Provides the OPV2V dataset, evaluation protocol, AP metric, and multi-agent frames that all weather variants and detection experiments are built on.","marker":"[7]"},{"why":"Motivates the parallel shared and collective fusion backbone design with identical sparse-convolution structure.","marker":"[29]"},{"why":"Supplies the PV-RCNN++ object detector used as the detection head on the fused features.","marker":"[30]"},{"why":"Cited as evidence that near-perfect LiDAR denoising in adverse weather is achievable on real and simulated data.","marker":"[31]"},{"why":"Cited as a comparable 3D point-cloud denoising model for adverse weather against which the denoising results can be viewed.","marker":"[32]"}],"fun_headline_variants":["Joint denoising and detection cuts LiDAR bandwidth by 23.6%","Denoise before sharing: 23.6% lower bandwidth, same detection AP","One LiDAR backbone filters rain, snow, fog and cuts bandwidth 23.6%","Shared network removes weather noise, keeps AP, cuts bandwidth 23.6%","Denoise before sharing: up to 23.6% less bandwidth for cooperative LiDAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated rain, snow, and fog point clouds from the authors' prior weather models, and the voxel labels they generate, are close enough to real-world LiDAR weather noise that a network trained on them will denoise real deployments equally well.","fun_headline_variants_meta":{"raw":{"variants":["Joint denoising and detection cuts LiDAR bandwidth by 23.6%","Denoise before sharing: 23.6% lower bandwidth, same detection AP","One LiDAR backbone filters rain, snow, fog and cuts bandwidth 23.6%","Shared network removes weather noise, keeps AP, cuts bandwidth 23.6%","Denoise before sharing: up to 23.6% less bandwidth for cooperative LiDAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3595,"prompt_tokens":1095,"completion_tokens":2500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":2390}},"tokens_in":711,"tokens_out":2500,"duration_ms":19257,"temperature":1.0,"reasoning_tokens":2390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:50:28.316257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DenoiseCP-Net on the simulated weather-augmented OPV2V data and evaluate it on real rainy, snowy, or foggy LiDAR sequences with physically validated noise labels; if per-voxel noise F1 drops substantially below the simulated near-99% levels, or if detection AP with denoised sharing falls below AP with noisy sharing, the central claim fails. A simpler check is to compute the transmitted payload size for real adverse-weather point clouds before and after denoising; if real noise does not inflate the voxel count, the reported 23.6% bandwidth reduction will not appear.","supporting_citations":[{"cited_title":"Simulating realistic rain, snow, and fog variations for comprehensive performance characterization of lidar perception,","cited_arxiv_id":null,"evidence_quote":"Supplies the rain, snow, and fog simulation models that generate the weather-augmented point clouds and the voxel-level noise ground-truth labels."},{"cited_title":"Mr3d-net: Dynamic multi-resolution 3d sparse voxel grid fusion for lidar-based collective perception,","cited_arxiv_id":null,"evidence_quote":"Establishes sparse voxel grids with coordinate-only sharing as the compact representation that DenoiseCP-Net transmits and denoises."},{"cited_title":"OPV2V: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,","cited_arxiv_id":null,"evidence_quote":"Provides the OPV2V dataset, evaluation protocol, AP metric, and multi-agent frames that all weather variants and detection experiments are built on."},{"cited_title":"S2s-net: Addressing the domain gap of heterogeneous sensor systems in lidar-based collective perception,","cited_arxiv_id":null,"evidence_quote":"Motivates the parallel shared and collective fusion backbone design with identical sparse-convolution structure."},{"cited_title":"PV-RCNN++: Point-voxel feature set abstraction with local vector representation for 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the PV-RCNN++ object detector used as the detection head on the fused features."},{"cited_title":"Cnn-based lidar point cloud de-noising in adverse weather,","cited_arxiv_id":null,"evidence_quote":"Cited as evidence that near-perfect LiDAR denoising in adverse weather is achievable on real and simulated data."}],"review_version":1}