{"id":"2480902a-4c23-4672-992f-1bedb81155f5","arxiv_id":"2506.16531","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CADC+ pairs each snowy CADC driving sequence with a matching clear-weather sequence from the same roads and period, enabling controlled evaluation of snow's impact on LiDAR 3D detection.","lead":"The authors release CADC+, a paired dataset of snowy and clear LiDAR driving scenes from the same roads and time period, and use it to test how snow affects 3D car detection. Snowy driving data is rare, so this resource lets researchers isolate weather effects instead of comparing scenes that differ in many ways.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pairing-quality confound is load-bearing: 21/74 sequences were manually matched to different locations, yet no quantitative pairing metric is reported; the snowy-vs-clear detection gap may partly be scene mismatch.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: pairing quality. I agree with the CONDITIONAL verdict because the dataset contribution is real and released, but the experimental attribution of the detection gap to snow depends on the unvalidated assumption that residual scene differences are small. The paper explicitly acknowledges imperfect pairing for 21 of 74 sequences and even for many 'matched' pairs, the matching required extending subsequences or accepting large lateral offsets. Without a quantitative report of per-pair coverage or distance, the central experimental claim is not fully supported. However, this is a condition to be addressed by additional analysis rather than a fatal flaw: the dataset itself is valuable, and the concern can be settled by measuring pairing quality and re-running the key comparisons on matched versus manual subsets. Therefore the reader's conditional recommendation stands unchanged.","tokens_in":11926,"tokens_out":4440,"duration_ms":44843,"concrete_test":"Compute, using the authors' cover(s,c,theta) and interpolation definitions (Appendices A-B), per-pair values of maximum lateral distance (d_max) and the fraction of snowy frames covered at theta thresholds, separately for the 53 closely matched and 21 manually matched pairs. Then rerun the Fig. 7 cross-evaluation (Base and SSL, L1/L2 AP) on the 53-pair subset and on the 21-pair subset, with at least 3 training seeds and reported standard deviations. If the snowy-vs-clear gap or the slope of snowy AP versus snowy training fraction changes materially between subsets, the attribution to snow alone is not supported; if both subsets show the same trends, the manual-pair confound is minor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim is that the performance gap between snowy and clear evaluation in Fig. 7 and Table II is caused by snow. That attribution requires the paired clear sequences to have negligible residual scene differences. The paper's own text undermines this: Section III-B says 'none of these requirements can be fully met'; Section III-D says 21 of 74 sequences could not be closely matched, with 15 manually paired to a different location (e.g., parking lot to parking lot) and 6 to a different road-agent composition; even Fig. 3d-e shows matched pairs extended beyond closest frames or with stationary segments to reach 100 frames, increasing maximum distance discrepancy. No quantitative pairing-quality metric is reported: no cover(s,c,theta) values at the chosen theta, no d_max distribution, no per-pair breakdown of closely matched versus manually matched pairs. The distribution comparisons in Fig. 6 show similar point/object counts but cannot control for spatially correlated road geometry or layout. Thus the observed roughly 3-4 AP gap and the trend in Fig. 7 may partly be scene mismatch rather than snow, weakening the aleatoric/epistemic uncertainty conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CADC+, an extension of the Canadian Adverse Driving Conditions (CADC) dataset with clear-weather sequences intended to match each of CADC's 74 snowy sequences. The stated goal is to provide the first paired real-data domain adaptation dataset for LiDAR-based 3D object detection in winter conditions, minimizing domain shift from factors other than snow. The authors describe a sequence-matching algorithm based on spatial interpolation and a coverage function, a sparse-labelling strategy with semi-supervised learning, and preliminary experiments comparing models trained on mixtures of snowy and clear data as well as on de-snowed synthetic clear data. They report that snow degrades detection performance and interpret this as evidence that snow introduces both aleatoric and epistemic uncertainty.","tokens_in":12142,"tokens_out":4434,"duration_ms":49461,"significance":"If the matching is as effective as claimed, CADC+ addresses a real gap: existing snowy driving datasets either lack paired clear data, lack sufficient labelled 3D boxes, or rely on synthetic de-snowing that introduces its own domain shift. The distribution comparisons in Fig. 6 (point counts per box, objects per frame) are useful evidence that the two domains are broadly similar in scene content. The de-snowing comparison in Table II is also a valuable negative result, showing that DROR and LiSnowNet do not produce realistic clear data under this evaluation protocol. The contribution is primarily a dataset, and the experimental section is explicitly preliminary, so the bar for the qualitative uncertainty claims should be adjusted accordingly. The paper's central weakness is that the pairing quality, which is load-bearing for attributing the observed performance gap to snow, is not quantified beyond qualitative descriptions.","major_comments":[{"comment":"The central attribution of the snowy-vs-clear performance gap to snow is not supported because pairing quality is never quantified. The paper states in Section III-B that \"none of these requirements can be fully met\" and in Section III-D that 21 of 74 CADC sequences could not be closely matched, with 15 manually paired to a different location and 6 to a different road-agent composition, yet no values of cover(s,c,theta) at the selected theta, no d_max distribution, and no per-pair breakdown of the matching subset are reported. Figure 3d-e also show that endpoint selection can extend matched sequences beyond the closest matching frames to reach 100 frames, increasing the maximum distance discrepancy. Without quantitative pairing statistics, the roughly 3-4 AP gap in Table II and the trends in Fig. 7 may partly reflect scene or layout mismatch rather than snow. Please report the coverage and distance statistics for all 74 pairs and repeat the analysis on the subset of closely matched pairs to show that the conclusions are robust.","section":"Section III-B and III-D, Fig. 7, Table II"},{"comment":"All experimental results are single-run, with no error bars, confidence intervals, or statistical tests. The aleatoric and epistemic uncertainty conclusions rest on small differences (for example, L1 AP gaps of 1-4 points in Table II and the non-monotonic peak at 75% snowy training data in Fig. 7), which could shift with random seed or training noise. Please provide multiple seeds or otherwise quantify variability, or temper the wording of the uncertainty conclusions so that they are presented as observations from a preliminary single-run study rather than established trends.","section":"Section IV, Fig. 7, Table II"}],"minor_comments":[{"comment":"The x-axis is labelled \"fraction of training data\" but the tick labels and the words \"snowy\" and \"clear\" suggest it is specifically the fraction of snowy training data; please state this explicitly in the caption.","section":"Figure 7"},{"comment":"The notation c*_{s,theta} is defined as an argmax and therefore denotes a set, but the surrounding prose sometimes refers to it in the singular; a brief clarification of set-valued notation would improve readability.","section":"Section III-B, Eqs. (1)-(3)"},{"comment":"The Delta% column reports differences in percentage points (e.g., -2.76), not relative percentages; please label the column as difference in percentage points or adjust the values.","section":"Table I"},{"comment":"The statement that training epochs are \"normalized with respect to the amount of data\" is imprecise; please report the actual epoch counts or effective iteration counts for each training configuration.","section":"Section IV"},{"comment":"The distribution comparisons in Figs. 5 and 6 are helpful, but the paper would benefit from explicitly stating whether the reported statistics include the 21 manually matched pairs, since those pairs may behave differently from the closely matched ones.","section":"Section III-D"}],"recommendation":"major_revision","confidential_remarks":"The dataset is likely to be a useful community resource, and the de-snowing evaluation is a sensible application. My recommendation of major_revision is driven by the need to quantify pairing quality and to acknowledge the single-run nature of the experiments; I do not see circularity or novelty issues with the dataset itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: CADC+ is a real, useful resource. It is the first paired clear/snow LiDAR driving dataset with 3D box annotations at this scale, built from the same platform and roads as CADC, and the authors did the hard work of matching 74 sequences and labeling the clear-side validation fully. The dataset itself deserves referee time.\n\nWhat is genuinely new is the pairing methodology and the artifact. The coverage-based matching with endpoint selection is a sensible extension of trajectory alignment, and the distribution comparisons (points per box, objects per frame, speeds) suggest the clear and snowy sets are broadly comparable, which is the best you can do with real data. The de-snowing comparison in Table II is a useful sanity check: it shows DROR and LiSnowNet do not produce data that behaves like real clear weather, which is a legitimate motivation for the dataset.\n\nThe soft spots are the experimental section and the pairing-quality check. Figure 7 and Table II are single runs without error bars or significance tests, so the aleatoric/epistemic distinction rests on trends that could move with noise. That is minor for a dataset paper, but it should be labeled as preliminary. More important is the confound the authors themselves admit in Sections III-B and III-D: 21 of 74 sequences were not closely matched, with 15 matched to a different location and 6 to a different road-agent composition. The paper reports no quantitative pairing metric, no cover values at the chosen theta, no d_max distribution, and no per-pair breakdown. Without that, part of the 3-4 AP gap between snowy and clear could be scene mismatch rather than snow. The authors' claims are not reckless; the distribution curves in Fig. 6 look close, and the gap is consistent across models. But the load-bearing claim that snow introduces aleatoric uncertainty depends on residual scene differences being negligible, and that is not shown.\n\nThe self-citation to their sparse-labelling work [25] is fine; the correctness of CADC+ does not depend on that result, and they validate SSL on CADC in Table I.\n\nWho is this for: anyone working on adverse-weather perception or domain adaptation for LiDAR. It is a dataset paper with preliminary experiments, not a definitive study of snow. With a supplemental pairing-quality analysis and multiple seeds, the experimental claims would be much stronger, but the dataset can be used now.\n\nRecommendation: send to peer review; the resource is worth refereeing, and the authors have been honest about the pairing limitations. Ask for a pairing-quality appendix and error bars.","headline":"CADC+ is a genuinely useful paired clear/snow LiDAR dataset and deserves peer review, but the snow-effect experiments need a pairing-quality analysis before the aleatoric/epistemic claim is solid.","tokens_in":12654,"tokens_out":2120,"would_cite":true,"duration_ms":20881,"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":"Snow is not just noise for LiDAR detection; it is a separate data domain, and CADC+ is the first paired dataset that measures both effects.","keywords":["paired weather dataset","LiDAR 3D object detection","domain adaptation","snowfall","Canadian Adverse Driving Conditions","semi-supervised learning","point cloud de-snowing","autonomous driving"],"falsifier":"Restrict evaluation to the 53 closely matched pairs and compute the clear-versus-snowy average-precision gap, then compute the same gap on the 21 manually matched pairs; if the loose pairs show a markedly larger gap, residual scene mismatch is contaminating the snow measurement, whereas equal gaps would support the pairing assumption. A second check: train a detector on real clear data and evaluate it on de-snowed snowy frames and on real clear frames from the same pairs; equal performance would show de-snowing is a valid substitute, which the paper predicts is false.","tokens_in":11723,"feed_emoji":"🌨️","tokens_out":10292,"duration_ms":95022,"temperature":0.7,"pith_summary":"This paper argues that previous LiDAR driving datasets cannot isolate the effect of snow on 3D object detection, because they either lack enough labelled snow and clear data or use synthetic clear data that introduces a second domain gap. To close that gap, it introduces CADC+, a paired extension of the CADC dataset in which each of 74 snowy sequences is matched to a clear sequence recorded on the same roads in the same period, with 3D bounding-box annotations. The paper's preliminary experiments indicate that snow degrades detection in two ways at once: as aleatoric noise that corrupts individual point clouds, and as an epistemic domain shift that makes snowy data unlike clear data for a trained detector. The dataset is designed so that a sympathetic reader can use it to separate these two effects and to test whether de-snowing or domain adaptation actually restores clear-weather performance.","feed_headline":"First paired clear-snow LiDAR dataset measures snow's 3D detection cost","feed_subtitle":"74 snowy drives matched to clear drives on the same roads isolate snow's effect from route and traffic differences.","key_machinery":"The central object is the paired dataset CADC+, built by matching each CADC snowy sequence to a clear sequence from the same road network and recording period. The mechanical core is a coverage-based matching procedure: a function measures what fraction of a snowy trajectory lies within a given lateral distance of a candidate clear trajectory, candidates are ranked by coverage quality, and endpoint sampling resamples the chosen clear sequence to at least 100 frames at 2 to 10 hertz. A second mechanism, sparse labelling with semi-supervised pseudo-labels, makes full-sequence training affordable while matching the performance of fully human-annotated training.","core_discovery":"The central claim is that snow is not a single failure mode for LiDAR-based 3D detection but a combination of two. One component is aleatoric: snow physically corrupts the measurement, adding noise, attenuation, and occlusion, so the same object is simply harder to sense. The other is epistemic: snowy point clouds form their own statistical domain, so a detector trained only on clear data performs worse on snow, and adding snowy training data monotonically improves snowy detection. The paper also finds that de-snowing real snowy clouds with a statistical filter (DROR) or a learned method (LiSnowNet) does not recreate clear weather from a detector's point of view: models trained on de-snowed data behave on clear evaluation data much like models trained on snow, which the authors take as evidence that synthetic clear data is not a substitute for real paired clear-snow data.","pith_inferences":["One consequence the authors do not draw explicitly: if snowy point clouds are a distinct domain, then weather-invariant detectors could be trained by feature-space alignment between the CADC+ pairs, using the paired geometry as supervision.","The 21 manually matched pairs form a natural stress test: a detector that is truly snow-robust should show the same clear-versus-snowy gap on close and loose pairs; divergence would mean route mismatch is contaminating the measurement.","The same matching protocol, coverage under lateral thresholds plus endpoint sampling, transfers to other seasonal or weather pairings whenever clear and adverse runs were collected on a shared route network.","The paper's own results suggest an operating point that it does not elevate into a rule: roughly balanced snowy/clear training appears to maximize worst-case performance across the two domains."],"forward_implications":["Detectors trained only on clear weather will systematically underperform on snow, and the gap cannot be closed by more clear data alone.","Training near half snowy and half clear data gives near-maximum performance on both domains, suggesting a practical recipe for all-weather detectors.","De-snowed point clouds are not a reliable substitute for real clear data when training or evaluating snow-effect studies; downstream detection performance, not point-cloud appearance, is the right yardstick.","Sparse labelling with semi-supervised learning makes a paired winter dataset feasible at roughly 10 percent annotation cost without losing detection accuracy.","CADC+ provides a controlled testbed for weather domain adaptation, letting future detectors be trained and evaluated on the same routes in both weathers."],"supporting_citations":[{"why":"Supplies the 74 labelled snowy sequences and their annotations that CADC+ extends with clear-weather matches.","marker":"[21]"},{"why":"Provides the sparse-labelling and semi-supervised auto-labelling method that justifies labelling only every tenth training frame.","marker":"[25]"},{"why":"Supplies the VoxelNeXt 5-frame-aggregated detector architecture used for all Base and SSL models.","marker":"[7]"},{"why":"Defines the L1 and L2 3D IoU-based average precision metrics used for all evaluations.","marker":"[24]"},{"why":"Supplies DROR, the statistical de-snowing baseline whose synthetic clear data is compared against CADC+ clear data.","marker":"[6]"},{"why":"Supplies LiSnowNet, the learned de-snowing baseline whose synthetic clear data is compared against CADC+ clear data.","marker":"[30]"},{"why":"Provides one existing labeled clear-and-snow dataset whose label counts the paper argues are too low for training modern detectors.","marker":"[4]"},{"why":"Provides another labeled clear-and-snow dataset whose clear and snow label counts the paper argues are too small.","marker":"[8]"}],"fun_headline_variants":["First paired snow-clear LiDAR dataset separates snow's noise from domain shift","Snow's 3D detection cost has two parts: noise and domain shift","De-snowing LiDAR doesn't recreate clear weather for 3D detectors","CADC+: the first paired snow-clear driving dataset isolates snow's effect","Snow introduces aleatoric noise and a distinct domain—CADC+ separates them"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that each paired clear and snowy drive is similar enough in route, layout, traffic, and timing that the remaining detection gap is caused by snow; the authors state in the matching section that none of the ideal pairing requirements can be fully met, and 21 of the 74 snowy sequences were matched manually to different locations or road layouts.","fun_headline_variants_meta":{"raw":{"variants":["First paired snow-clear LiDAR dataset separates snow's noise from domain shift","Snow's 3D detection cost has two parts: noise and domain shift","De-snowing LiDAR doesn't recreate clear weather for 3D detectors","CADC+: the first paired snow-clear driving dataset isolates snow's effect","Snow introduces aleatoric noise and a distinct domain—CADC+ separates them"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2541,"prompt_tokens":949,"completion_tokens":1592,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1489}},"tokens_in":565,"tokens_out":1592,"duration_ms":12523,"temperature":1.0,"reasoning_tokens":1489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:23:48.287706+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Restrict evaluation to the 53 closely matched pairs and compute the clear-versus-snowy average-precision gap, then compute the same gap on the 21 manually matched pairs; if the loose pairs show a markedly larger gap, residual scene mismatch is contaminating the snow measurement, whereas equal gaps would support the pairing assumption. A second check: train a detector on real clear data and evaluate it on de-snowed snowy frames and on real clear frames from the same pairs; equal performance would show de-snowing is a valid substitute, which the paper predicts is false.","supporting_citations":[{"cited_title":"Canadian adverse driving conditions dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the 74 labelled snowy sequences and their annotations that CADC+ extends with clear-weather matches."},{"cited_title":"3D object detection with track-based auto-labelling using very sparsely labelled data,","cited_arxiv_id":null,"evidence_quote":"Provides the sparse-labelling and semi-supervised auto-labelling method that justifies labelling only every tenth training frame."},{"cited_title":"V oxelNeXt: Fully sparse V oxelNet for 3D object detection and tracking,","cited_arxiv_id":null,"evidence_quote":"Supplies the VoxelNeXt 5-frame-aggregated detector architecture used for all Base and SSL models."},{"cited_title":"Scalability in perception for au- tonomous driving: Waymo Open Dataset,","cited_arxiv_id":null,"evidence_quote":"Defines the L1 and L2 3D IoU-based average precision metrics used for all evaluations."},{"cited_title":"De-noising of lidar point clouds corrupted by snowfall,","cited_arxiv_id":null,"evidence_quote":"Supplies DROR, the statistical de-snowing baseline whose synthetic clear data is compared against CADC+ clear data."},{"cited_title":"LiSnowNet: Real-time snow removal for LiDAR point clouds,","cited_arxiv_id":null,"evidence_quote":"Supplies LiSnowNet, the learned de-snowing baseline whose synthetic clear data is compared against CADC+ clear data."},{"cited_title":"Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather,","cited_arxiv_id":null,"evidence_quote":"Provides one existing labeled clear-and-snow dataset whose label counts the paper argues are too low for training modern detectors."},{"cited_title":"Ithaca365: Dataset and driving perception under repeated and challenging weather conditions,","cited_arxiv_id":null,"evidence_quote":"Provides another labeled clear-and-snow dataset whose clear and snow label counts the paper argues are too small."}],"review_version":2}