{"id":"1ae62c3c-e2e0-41a9-9e48-57af9c1d0415","arxiv_id":"2508.13485","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A LiDAR-supervised voxel-mask loss improves 4D radar-only 3D detection by teaching the network to suppress noise, with noted experimental ambiguities.","lead":"CORENet uses LiDAR point clouds during training to generate a per-voxel validity mask that teaches a radar-only detector which radar points are real, and then runs on radar alone at inference. The authors report large gains on the Dual-Radar dataset, but unclear data selection and inconsistent tables leave the result conditional.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main results and ablations report incompatible APs for the same CasA-V+CORENet configuration, so the headline improvement numbers are not internally consistent.","rationale":"I agree with the reader that the paper deserves a conditional verdict, but I locate the load-bearing concern differently. The reader's weakest assumption about the LiDAR-distance mask (Eq. 4-5) is a legitimate physical concern, yet even if that assumption is wrong, the reported gains are empirical facts on this dataset; the mask issue matters for transferability, not for whether the Table I numbers support the claim. By contrast, the discrepancy between Table I and Tables II–IV directly attacks the central claim: the same configuration yields 49.46 and 67.34 Car 3D AP. This is not a minor typo; it is an 18-point gap. Without an explanation (e.g., validation vs test split, different difficulty level, or different IoU), the reader cannot tell which number is the real result, and the claimed +34.64 improvement may be either understated or an artifact of an undocumented split. The frame-selection issue is also concerning but secondary. A single reproduction of the complete model on the test split would resolve which set of numbers is correct. I therefore keep the verdict CONDITIONAL (unchanged from the reader), because the paper needs this clarification/reproduction before its central claim can be accepted.","tokens_in":10773,"tokens_out":5680,"duration_ms":56925,"concrete_test":"Run the released CORENet code (linked in the paper) with the complete HMSD-Net, tau=0.5, single-layer MLP, and train on the stated 5100-frame training split for 80 epochs. Evaluate on the same test split used for Table I (the 2400-frame test set). If the resulting Car 3D AP is ≈49.46, Table I is confirmed and Tables II–IV must be on a different split; if it is ≈67.34, Table I is contradicted. Additionally, ask the authors to state explicitly which split (train/test/val) is used in Tables II–IV and what 'effective selection' of 9966 frames removed from the Dual-Radar set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I gives CasA-V+CORENet Car 3D AP 49.46 and BEV 54.38. Tables II–IV, described as ablations of the same integrated model, give the identical configuration (HPNet+SConvNet, tau=0.5, 1-layer MLP) Car 3D AP 67.34 and BEV 71.23. These are ~18 AP points apart with no stated difference in dataset split, IoU, or evaluation protocol. The central claim—that CORENet adds +34.64/+37.51 over the CasA-V baseline—is therefore not backed by a consistent set of numbers. If 49.46 is the test-set value, the ablations must have been run on another split and cannot be compared to Table I; if 67.34 is the test-set value, the reported gains are actually larger but Table I is wrong. Either way, the paper does not provide a trustworthy basis for the claimed improvement. The unexplained selection of 9966 frames compounds this: without criteria, it is unclear whether the test split is representative.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CORENet, a plug-and-play cross-modal denoising module for 4D radar voxel-based 3D object detection. During training, LiDAR points are used to generate binary supervision masks by thresholding nearest-neighbor distances (KDTree, tau = 0.5 m); the radar stream passes through a hierarchical multi-scale denoising network (HMSD-Net, consisting of HPNet and SConvNet) and a Noise Predictor whose output is compared with the mask via Smooth L1 loss, added to the detection losses. At inference, LiDAR is not needed. The method is evaluated on a selected subset of the Dual-Radar dataset (9966 frames), and Table I reports large gains when CORENet is integrated with CasA-V and Voxel R-CNN.","tokens_in":11095,"tokens_out":6057,"duration_ms":66950,"significance":"The inference-time radar-only property and the plug-and-play design are practically attractive, and the paper makes a reasonable step toward using LiDAR supervision to denoise 4D radar. The method's modularity and the inclusion of a code repository are strengths. However, the strength of the evidence is currently compromised by a large internal inconsistency between the test-table results and the ablation tables for the same configuration, by an unexplained dataset filtering procedure, and by the omission of the closest prior method (RadarDistill). The core idea is not entirely new, but the specific noise-mask formulation and hierarchical denoising architecture are reasonable contributions. If the inconsistencies are resolved and the results are reproducible, the paper would be a useful contribution to the radar perception community.","major_comments":[{"comment":"The same configuration (CasA-V + CORENet, HPNet + SConvNet, KDTree distance 0.5 m, 1-layer Noise Predictor) reports Car 3D AP 49.46% and BEV AP 54.38% in Table I, but Car 3D AP 67.34% and BEV AP 71.23% in Table II (also rows in Tables III and IV). These values differ by roughly 18 points, with no stated difference in dataset split, IoU threshold, or evaluation protocol. The mean AP differs similarly (27.47% vs 21.37% 3D; 29.01% vs 23.14% BEV). This directly undermines the headline improvement claim of +34.64%/+37.51% over the CasA-V baseline, because a reader cannot determine which set of numbers constitutes the actual test result. The authors must clarify the split used for each table and either rerun the ablations on the test set or explain the discrepancy.","section":"Sec. IV-C vs. Sec. IV-D, Tables I and II-IV"},{"comment":"The dataset description states 'After effective selection, the experimental dataset comprises 9966 frames' but gives no criteria for this selection. The Dual-Radar dataset composition (sequences, weather, traffic conditions) is not described, and the choice of 5100/2400/2466 train/test/validation frames is not justified. If the selection favors easy scenes or particular sensor configurations, the reported gains may not generalize. The authors need to specify the selection protocol, ideally with a random or stratified split and a comparison with the full dataset statistics.","section":"Sec. IV-A"},{"comment":"RadarDistill [16] is cited as a LiDAR-to-radar distillation method, which is the closest prior art to the paper's approach of using LiDAR supervision to improve radar-only detection. It is not included in any quantitative comparison. Since the paper claims 'superior performance compared to existing mainstream approaches,' a direct comparison with RadarDistill (preferably with the same detector and data split) is necessary to isolate the benefit of CORENet's specific mask-based supervision versus generic feature distillation.","section":"Sec. II and Sec. IV-C, Table I"},{"comment":"The experimental setup says the model is trained for 80 epochs and 'we presented the best results obtained.' No standard deviation, multiple seeds, or repeated runs are reported. Given the unusually large reported gains and the AP inconsistency between tables, the reader cannot assess the stability of the results. The authors should report mean and standard deviation over at least three runs, or otherwise justify why a single run is sufficient.","section":"Sec. IV-B"},{"comment":"The binary supervision mask labels any radar point with no LiDAR point within tau = 0.5 m as invalid. This equates 'no LiDAR neighbor within tau' with 'noise,' which may be incorrect for radar points that originate from genuine objects that LiDAR does not sample at that distance, or from multi-path reflections that carry structural information. No analysis is provided of how many points are suppressed by this mask or whether ground-truth object points are systematically retained. The validation-based choice of tau mitigates the issue to some extent, but the core premise deserves empirical scrutiny, e.g., showing precision/recall of the mask against manually or automatically labeled valid radar points.","section":"Sec. III-B, Eq. (4)-(5)"}],"minor_comments":[{"comment":"Equation (3) defines the classification loss as binary cross-entropy (BCE), while Section III-C states the classification loss uses focal loss. Please align the notation and the description.","section":"Sec. III-A, Eq. (3), Sec. III-C"},{"comment":"The text says 'nearest-neighbor search with a fixed distance' but Eq. (4) defines a radius search over all points within a threshold. Clarify whether the mask uses the nearest neighbor only (as the text suggests) or all points within tau (as the equation indicates).","section":"Sec. III-B, Eq. (4)"},{"comment":"Tables II-IV use checkmark symbols to denote whether a component is used. These may not render in all formats. Consider using 'Yes'/'No' or explicit model names.","section":"Sec. IV-D, Tables II-IV"},{"comment":"The introduction states the method 'improves the mean accuracy of the original model by 7%-18% in 3D and BEV views.' This is vague; specify whether this refers to mean AP over categories, the Car-only AP, and the specific integration models.","section":"Sec. I"},{"comment":"The sentence 'Voxel convolution can capture local spatial information but fails to preserve topological relationships between points' is not elaborated. Define what 'topological relationships' means in this context and how HPNet captures them.","section":"Sec. III-A"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between Table I and Tables II-IV is the most serious issue; it must be resolved before the paper can be considered reliable. Please ask the authors to clearly state the evaluation split for each table, rerun the ablations on the same test set, and report variance. Also require a detailed dataset selection protocol and a RadarDistill baseline. If these cannot be provided, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CORENet's core idea—train a voxel-level validity mask from LiDAR proximity and use it as an auxiliary loss in a tandem denoising network—is a legitimate variation on RadarDistill, and the write-up is mostly clear. But the headline numbers don't survive cross-checking. Table I says CasA-V+CORENet gets 49.46% Car 3D AP on the test set; Tables II–IV, described as ablations of that same integration, report 67.34% for the \"complete\" configuration. That's an 18-point gap with no stated split difference. Either the ablations ran on validation and the paper never says so, or the main table is wrong. Either way, the +34.64 claim is not backed by internally consistent evidence.\n\nWhat's new: a per-voxel binary mask supervised by a KDTree nearest-neighbor test against LiDAR, combined with a hierarchical PointNet + sparse-conv denoiser, with radar-only inference. That's not the same as feature-level distillation in RadarDistill, and the idea is worth exploring. The method section is readable, the loss is standard, and the authors honestly note they still need LiDAR at training time and that small-object performance is weak.\n\nSoft spots, in rough order of seriousness. First, the table inconsistency above. Second, RadarDistill—the closest prior work—is cited but never compared. That's a hard omission for a \"cross-modal supervision\" paper. Third, the \"effective selection\" of 9966 frames from Dual-Radar has no criteria; without that, the test split's representativeness is unknown. Fourth, no error bars or multiple seeds; many reported differences (e.g., 0.3 vs 0.5 m KDTree) are within plausible noise. Fifth, the intro claims 7–18% mean AP gains, but the table shows 16+ points absolute on mean 3D AP and 34+ on car, so the abstract/intro understate or misstate their own results. Finally, the mask target assumes \"no LiDAR within 0.5 m\" means noise; that's reasonable as a heuristic but not defended, and it could suppress legitimate radar detections that LiDAR misses.\n\nWho this is for: researchers working on radar-only detection and cross-modal distillation. The idea has legs, but the current manuscript is not a trustworthy basis for the claimed improvement.\n\nRecommendation: send it to peer review, but expect a heavy revision request. The reviewer must ask for a reconciled set of numbers (same split, same protocol), a RadarDistill comparison, and selection criteria. If those come back, the paper could be worth keeping; as it stands, I wouldn't cite it.","headline":"Plausible cross-modal denoising idea, but the reported numbers don't cross-check: Table I and the ablations differ by ~18 AP points for the same configuration, so the headline gain is not trustworthy as written.","tokens_in":11523,"tokens_out":2803,"would_cite":false,"duration_ms":28888,"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":"This paper claims that a plug-in denoising module trained with a LiDAR-derived voxel mask can lift radar-only 3D car detection on the Dual-Radar dataset by over 34 percentage points, and into double digits for small objects.","keywords":["4D radar","radar point cloud denoising","LiDAR supervision","cross-modal learning","3D object detection","voxel-based detection","autonomous driving","Dual-Radar dataset"],"falsifier":"Train the identical CORENet on the Dual-Radar frames with the LiDAR-derived mask replaced by a mask shifted by 2 m in a random direction; if car 3D AP stays near 49%, the mask loss is not the source of the gain. Alternatively, evaluate the trained model on objects visible to radar but beyond LiDAR range, or in fog where LiDAR drops out: if detection on those objects collapses, the 0.5 m rule is discarding genuine radar returns.","tokens_in":10737,"feed_emoji":"📡","tokens_out":5499,"duration_ms":55333,"temperature":0.7,"pith_summary":"This paper claims that 4D radar detection can be materially improved by teaching a plug-in denoising module what noise looks like, using LiDAR as a teacher during training only. The proposed CORENet adds a hierarchical multi-scale denoising network (HMSD-Net) and a voxel-level mask loss computed by matching each radar voxel to LiDAR points within 0.5 m. Integrated with the CasA-V detector on the Dual-Radar dataset, it raises car 3D AP from 14.81% to 49.46% and BEV AP from 16.87% to 54.38%, with mean AP across car, cyclist, and pedestrian also improving over classical denoising baselines. If this holds, radar-only perception can absorb LiDAR-level supervision without needing LiDAR at inference, which matters for robust autonomous driving in weather where LiDAR degrades.","feed_headline":"LiDAR-trained voxel mask lifts radar car detection by 34.6 AP points","feed_subtitle":"CORENet uses LiDAR only during training to denoise 4D radar, adding a voxel-mask loss to existing detectors.","key_machinery":"The central object is the voxel-validity mask: a binary label per radar voxel, produced by a KDTree nearest-neighbor search that marks a radar voxel valid if any LiDAR point lies within 0.5 m (Eqs. 4–5). HMSD-Net predicts this mask from radar data alone, and the Smooth L1 voxel-mask loss (weight 50) pulls the network toward keeping LiDAR-consistent voxels while discarding sidelobe noise. The mask is the conduit through which LiDAR knowledge enters the radar-only detector at training time.","core_discovery":"The central claim is that cross-modal supervision — a binary mask derived from LiDAR proximity plus a denoising head appended to an existing voxel-based detector — lets a radar-only network suppress the sidelobe noise that dominates 4D radar point clouds. HMSD-Net combines point-topology features (HPNet) with sparse-convolution spatial features (SConvNet) and predicts a per-voxel validity score; the score is trained against the LiDAR-derived mask with Smooth L1 loss. At inference the mask generator and LiDAR are dropped. On the Dual-Radar test set, CasA-V + CORENet reaches 49.46% 3D AP and 54.38% BEV AP for cars, absolute gains of +34.64 and +37.51 points over CasA-V, and the Voxel R-CNN var","pith_inferences":["A testable extension not pursued in the paper is to replace the LiDAR-derived mask with a shifted or corrupted mask: if AP stays high, the mask loss is not the active ingredient; if AP collapses, the 0.5 m neighbor rule is doing the work.","The 0.5 m threshold ties the method's ceiling to LiDAR sampling density; on long-range or occluded objects where LiDAR returns are absent, the supervision would label genuine radar points as noise, so gains may shrink exactly where radar is most valuable.","The same denoising-by-mask idea could be trained with camera-derived depth or occupancy pseudo-labels, removing even the training-time LiDAR requirement, at the cost of a weaker teacher.","The paper reports gains on a selected subset of Dual-Radar frames without stating the selection criteria, so an independent evaluation on the full dataset would clarify how much of the improvement is due to the method rather than the sample."],"forward_implications":["Radar-only detectors can benefit from LiDAR-quality supervision without any LiDAR input at inference, preserving cost and robustness advantages of radar.","The plug-in design means the same denoising module and mask loss can be attached to any voxel-based 3D detector, not just CasA-V and Voxel R-CNN.","Classical denoising filters (ROR, SOR) are outperformed by the learned denoiser, suggesting that noise in 4D radar is structured enough to be learned rather than removed by statistics alone.","Small-object detection (pedestrians and cyclists) improves more in relative terms than car detection, indicating the denoiser helps exactly the sparse, few-point objects that typically fail.","The cross-modal supervision could be extended to other tasks built on radar voxels, such as segmentation or tracking, by reusing the same mask target."],"supporting_citations":[{"why":"Supplies the Dual-Radar dataset with Arbe Phoenix 4D radar and LiDAR data used for all experiments.","marker":"[4]"},{"why":"Defines CasA-V, the baseline detector whose +34.64% 3D AP and +37.51% BEV AP gains are the paper's main result.","marker":"[40]"},{"why":"Supplies Voxel R-CNN, the second voxel-based detector used to demonstrate that CORENet transfers across architectures.","marker":"[34]"},{"why":"Provides PointPillars, one of the baseline detection models compared in Table I.","marker":"[38]"},{"why":"Provides the ROR radius-outlier-removal denoising baseline that CORENet is compared against.","marker":"[10]"},{"why":"Provides the SOR statistical-outlier-removal denoising baseline that CORENet is compared against.","marker":"[11]"},{"why":"Motivates the cross-modal distillation idea of using LiDAR features to supervise radar representations.","marker":"[16]"},{"why":"Supplies the radar-imaging principle that sidelobe energy leakage around strong targets is the dominant noise source.","marker":"[12]"},{"why":"Provides the submanifold sparse convolution operation used as the building block of SConvNet.","marker":"[36]"},{"why":"Supplies the Smooth L1 loss used for both the box regression and the voxel-mask supervision.","marker":"[37]"}],"fun_headline_variants":["CORENet: LiDAR-trained radar denoiser lifts car detection by 34.6 AP","Radar-only inference, LiDAR-trained denoising, +34.6 AP for cars","Plug-and-play mask from LiDAR boosts radar car AP by 34.6","CORENet: radar-only detection, LiDAR-trained denoising, +34.6 AP","Plug-and-play radar denoiser uses LiDAR supervision to gain 34.6 AP"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that a radar voxel is valid exactly when a LiDAR point sits within 0.5 m of it, and that the 9,966 selected Dual-Radar frames fairly represent noisy 4D radar; if the threshold mislabels genuine radar returns or the frame selection is skewed, the reported AP gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["CORENet: LiDAR-trained radar denoiser lifts car detection by 34.6 AP","Radar-only inference, LiDAR-trained denoising, +34.6 AP for cars","Plug-and-play mask from LiDAR boosts radar car AP by 34.6","CORENet: radar-only detection, LiDAR-trained denoising, +34.6 AP","Plug-and-play radar denoiser uses LiDAR supervision to gain 34.6 AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001089,"raw_usage":{"total_tokens":4385,"prompt_tokens":739,"completion_tokens":3646,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":3527}},"tokens_in":483,"tokens_out":3646,"duration_ms":26285,"temperature":1.0,"reasoning_tokens":3527,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:00:09.824419+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical CORENet on the Dual-Radar frames with the LiDAR-derived mask replaced by a mask shifted by 2 m in a random direction; if car 3D AP stays near 49%, the mask loss is not the source of the gain. Alternatively, evaluate the trained model on objects visible to radar but beyond LiDAR range, or in fog where LiDAR drops out: if detection on those objects collapses, the 0.5 m rule is discarding genuine radar returns.","supporting_citations":[{"cited_title":"Dual Radar: A Multi-modal Dataset with Dual 4D Radar for Autonomous Driving","cited_arxiv_id":"2310.07602","evidence_quote":"Supplies the Dual-Radar dataset with Arbe Phoenix 4D radar and LiDAR data used for all experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines CasA-V, the baseline detector whose +34.64% 3D AP and +37.51% BEV AP gains are the paper's main result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Voxel R-CNN, the second voxel-based detector used to demonstrate that CORENet transfers across architectures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides PointPillars, one of the baseline detection models compared in Table I."},{"cited_title":"Shamsfakhr, D","cited_arxiv_id":null,"evidence_quote":"Provides the ROR radius-outlier-removal denoising baseline that CORENet is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SOR statistical-outlier-removal denoising baseline that CORENet is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the cross-modal distillation idea of using LiDAR features to supervise radar representations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the radar-imaging principle that sidelobe energy leakage around strong targets is the dominant noise source."},{"cited_title":"Graham, M","cited_arxiv_id":null,"evidence_quote":"Provides the submanifold sparse convolution operation used as the building block of SConvNet."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Smooth L1 loss used for both the box regression and the voxel-mask supervision."}],"review_version":1}