ModelNet40-E: An Uncertainty-Aware Benchmark for Point Cloud Classification

Pedro Alonso, Chongshou Li, Tianrui Li

Keywords: Point Cloud Classification, Robustness Benchmark, Sensor Noise, Uncertainty Estimation, Model Calibration

Abstract: Most existing benchmarks for point cloud classification focus solely on accuracy, overlooking critical aspects such as calibration and uncertainty awareness that are essential for safety-critical applications. We introduce ModelNet40-E, a benchmark that complements prior robustness efforts by providing noise-corrupted point clouds together with per-point uncertainty annotations via Gaussian parameters ($\sigma$, $\mu$). Unlike benchmarks based on random corruptions, ModelNet40-E introduces physically motivated LiDAR-like noise at multiple levels, reflecting real-world sensing conditions. Using this benchmark, we evaluate a range of representative point cloud architectures across varying noise levels. Our results show that accuracy alone can be misleading: some models with lower accuracy exhibit better calibration and uncertainty awareness, underscoring the need to evaluate all three metrics together.

Primary Area: datasets and benchmarks

Submission Number: 23249