Beyond the Snowflakes: Disentangling Weather Effects from Confounding Meta-Factors in LiDAR Object Detection

I. INTRODUCTION

• Snowflakes can obscure objects and create transient noise points, with errors potentially propagating to safety-critical downstream modules.Adverse weather can also introduce occlusion and false returns.Source: §I · p. 1
• Snow research remains comparatively limited, with scarce systematic quantitative analysis across snowfall intensities because data collection is difficult and costly.Prior studies often rely on theoretical scattering models or qualitative observations.Source: §I · p. 1
• The authors introduce WADS-3D and evaluate it alongside CADC, finding detection performance can improve with heavier snowfall as object complexity outweighs environmental complexity.Ablation, error-matrix, and meta-information analyses attribute the trend to confounding variables rather than falling snow directly.Source: §I · p. 1

A. LiDAR-based 3D Object Detection

• Modern LiDAR detectors use voxel, pillar, point-based, or hybrid representations; this study focuses on LiDAR-only methods to isolate snowy-condition effects.The literature also includes geometry-based clustering and sensor-fusion approaches.Source: A. LiDAR-based 3D Object Detection · p. 2

B. Impact of Snow on LiDAR Perception

• Prior snow studies report degradation from airborne effects, while heterogeneous sensor datasets can introduce domain shifts that complicate interpretation of weather effects.Simulation work often models airborne snow without surface accumulation; other work reports vehicle-detection drops of up to 50% under heavy snow.Source: B. Impact of Snow on LiDAR Perception · p. 2

C. Object- and Scene-Level Complexity Metrics

• Object- and scene-level complexity measures vary across benchmarks, including object point counts, visibility, range, annotation effort, geometric measures, and point-cloud entropy.The cited examples include LiDAR-centric object metrics in Waymo and post-hoc error analysis in nuScenes,.Source: C. Object- and Scene-Level Complexity… · p. 2

D. Analyzing the Influence of Meta-Information

• Meta-information studies have used Random Forest and SHAP for camera detectors and Kendall’s Tau and SHAP for LiDAR pedestrian detection.Prior work also examined false-negative and false-positive errors to account for multivariate dependencies [47].Source: D. Analyzing the Influence of… · p. 2

E. Research Gap

• Existing snowy-LiDAR studies often use coarse weather categories, omit surface accumulation, or combine different sensors, complicating interpretation of weather effects.The paper proposes independent WADS-3D and CADC benchmarks and meta-information analysis to address these gaps.Source: E. Research Gap · p. 2

A. Dataset

• WADS-3D contains 4,100 labeled frames and 30,104 cars; CADC contributes 7,000 LiDAR frames captured with a different sensor suite.WADS-3D is derived from WADS and uses 2,050 unique 360-degree scans rotated for forward-view requirements.Source: A. Dataset · p. 2
• WADS-3D vehicles have higher average points per bounding box despite hardware differences, and are generally observed farther away with lower mean intensity.Car dimensions also vary across the datasets.Source: A. Dataset · p. 2
• Figure 2 compares car distance, points per car, mean intensity, and dimensions across KITTI, CADC, and WADS-3D; WADS-3D has the highest points-per-car distribution.The caption attributes CADC’s lower point count to its 32-channel lidar.Source: Figure 2 · p. 3

B. Snowfall Intensity Classification

• Snowfall bins use snow-point counts from DSOR and LIOR filtering, with quantile binning to keep comparable numbers of ground-truth cars per bin.Snow points range from 764–19,386 (0.69%–28.34%, avg. 8.97%) in WADS-3D and 142–4,447 (0.66%–21.38%, avg. 6.71%) in CADC.Source: B. Snowfall Intensity Classification · p. 3

C. Evaluation Metircs

• The evaluation uses KITTI’s 3D detection protocol and Average Precision to summarize the precision-recall curve.The same metric is used for the ablation study to enable comparison with the main results.Source: C. Evaluation Metircs · p. 3

D. Implementation Details

• Five detectors span voxel, pillar, point-based, and hybrid architectures; WADS-3D uses KITTI-pretrained models, while CADC uses a 50:50 fine-tuning and evaluation split.CADC training data is snow-filtered, but evaluation uses original snowy frames.Source: D. Implementation Details · p. 3

E. Analysis Methodology

• The framework uses raw-versus-denoised ablations, error matrices, and meta-information rankings to analyze snowfall effects and detector outcomes.It applies Kendall’s Tau, Mutual Information, and Random Forest with SHAP analysis.Source: E. Analysis Methodology · p. 3
• Meta-information is divided into object-level features, including point count and distance, and scene-level features, including snowfall intensity and spatial complexity.The Random Forest models predict false negatives and true positives; conditional sampling is used for SHAP to limit extrapolation with correlated features.Source: E. Analysis Methodology · p. 4

1) Results on WADS-3D:

• On WADS-3D, AP rises with snowfall intensity across architectures, and the positive trend remains after snow filtering.Filtering improves absolute scores, but does not remove the trend.Source: 1) Results on WADS-3D · p. 4

2) Results on CADC:

• CADC confirms a positive AP trend across snowfall intensities, but with greater volatility than WADS-3D.Snow filtering preserves the positive trend without improving CADC performance.Source: 2) Results on CADC · p. 4, Figure 4 · p. 4

3) Impact of Explicit Snow Filtering:

• Snow filtering improves WADS-3D AP by an average of 6.72%, but yields negligible gains on CADC.The authors associate WADS-3D’s greater snow-noise volume with clutter and object concealment, unlike CADC’s lower number of snow points.Source: 3) Impact of Explicit Snow Filtering · p. 4, Figure 4 · p. 4

4) Architectural Robustness Under Snowfall:

• Voxel-based SECOND-IoU achieves the highest performance, ahead of pillar-based PointPillars, hybrid PV-RCNN, and point-based Part-A2.The authors suggest voxel aggregation filters sparse noise and that anchor-based models outperform anchor-free Part-A2.Source: 4) Architectural Robustness Under… · p. 4

• In WADS-3D, increasing snowfall intensity coincides with more true positives, fewer false negatives, and PR curves shifting upward and rightward despite background-clutter fluctuations.Background errors dominate localization errors among false positives.Source: 1) Error Decomposition &… · p. 4, Figure 5 · p. 5
• In CADC, false negatives decline across snowfall levels, while PR curves shift outward relative to WADS-3D.The outward shift delays precision dropoff and maintains high precision in the high-recall regime.Source: 1) Error Decomposition &… · p. 4, Figure 6 · p. 5
• Across the reported results, increased true positives and reduced false negatives outweigh increased background clutter, driving the net AP increase.Source: 1) Error Decomposition &… · p. 5

2) The Physical Paradox:

• Although snow noise is expected to attenuate signals and occlude objects, objects are easier to detect in heavy-snow frames, suggesting object characteristics may explain the improvement.Source: 2) The Physical Paradox · p. 5

3) Transition to Meta-Analysis:

• The authors analyze recall through a binary true-positive/false-negative framework to test whether heavy-snow subsets contain easier-to-detect objects.The analysis considers physical object properties such as distance, point count, and size.Source: 3) Transition to Meta-Analysis · p. 5

1) Analysis of Recall Determinants on WADS-3D:

• Across five architectures, object point count and distance are the strongest recall factors; snowfall intensity has negligible feature importance (<0.03).Point count and distance have Kendall’s τ of approximately 0.3 and −0.3, respectively; together they receive about 40–50% of model decision weight.Source: 1) Analysis of Recall Determinants on… · p. 5, Figure 7 · p. 6

2) Analysis of Recall Determinants on CADC:

• On CADC, object point count is the primary recall determinant (SHAP ≈0.55); mean point intensity is secondary (Kendall’s τ ≈0.3; SHAP ≈0.18), while snowfall intensity is negligible.Object dimensions also have negligible impact, likely because larger vehicles do not yield more returns at the given sensor resolution.Source: 2) Analysis of Recall Determinants on… · p. 6

D. Identifying Key Meta-Factors Driving Recall

• Across snowfall bins, object point count correlates positively with severity in WADS-3D (r = 0.820, p < 0.01) and CADC (r = 0.733); CADC mean object intensity also correlates positively (r = 0.694).The analysis uses Pearson correlations between snowfall intensity bins and dataset-specific recall drivers; WADS-3D ego-distance shows no significant trend (r = −0.282, p = 0.43).Source: D. Identifying Key Meta-Factors… · p. 6, Table III · p. 6

E. Discussion and Hypotheses

• The discussion proposes scene smoothing, snow returns entering car boxes, and a specular-to-diffuse reflectance shift as possible explanations for snow-related detection gains.The authors state that the physical mechanisms behind the statistical role of object point count remain uncharacterized.Source: E. Discussion and Hypotheses · p. 6
• Controlled findings support two physical changes from snow accumulation: geometric inflation that raises point count and a radiometric shift that enhances return intensity.The reported intensity enhancement is especially apparent at high incidence angles and long ranges.Source: E. Discussion and Hypotheses · p. 7

A. Conclusion

• Across WADS-3D and CADC, heavier snowfall coincides with better 3D detection because increased object point count outweighs falling-snow noise, making weather-bin-only robustness evaluation incomplete.The conclusion calls for safety validation to disentangle precipitation noise from correlated geometric changes.Source: A. Conclusion · p. 7

B. Future Work

• Future work will test whether accumulation similarly enhances geometry for smaller, non-rigid targets or whether atmospheric attenuation dominates their smaller cross-sections.The authors note that this may require targeted data collection.Source: B. Future Work · p. 7
• Planned follow-ups include dynamic point-density normalization, camera-LiDAR fusion, and simulations that model physical snow accumulation on scene geometry.These mitigation strategies target geometric distortions introduced by snow accumulation; the simulation direction moves beyond airborne-noise-only models.Source: B. Future Work · p. 7