A preprint posted on arXiv on 9 October by Michigan Technological University researchers argues that evaluating 3D LiDAR object detectors only by snowfall-severity categories can be misleading. The team compared five detection architectures on the snowy-driving datasets WADS-3D and CADC. WADS-3D contains 4,100 labeled frames generated from 2,050 unique LiDAR scans and 30,104 labeled cars; CADC contributes 7,000 frames captured with a different sensor. In both datasets, average precision rose as snowfall intensity increased.

Snow particles add noise, but the researchers suggest that cars also appear with more LiDAR points during heavier snow, potentially offsetting that noise; they do not establish this as a causal effect. Snow filtering increased average precision by a reported 6.72% on WADS-3D, while gains on CADC were negligible. Differences across the datasets point to the need to assess object distance, point density and sensor characteristics alongside weather severity. The study does not measure road safety or real-vehicle driving performance; it is a preprint that has not yet undergone peer review.