Robust Deep 3D Point Cloud Analysis

Abstract

3D point clouds are a fundamental representation for geometric perception and support applications such as autonomous driving, robotics, and 3D scene understanding. In practice, point clouds are captured under imperfect sensing conditions, suffering from measurement noise, occlusions, density variations, and arbitrary object orientations, which can severely compromise the reliability of recognition models. Despite rapid progress in deep learning for point cloud analysis, many methods are still trained and evaluated on clean, well-aligned data and thus degrade sharply under corruptions or unseen rotations. This thesis systematically studies and improves the robustness of point cloud recognition against both natural corruptions and geometric transformations. We first conduct an empirical robustness analysis on ModelNet40-C to characterize vulnerabilities across corruption types. Guided by these findings, we propose techniques including spatial sorting and a set-mixing aggregation module to stabilize local features under structural noise. We further tackle rotation robustness by introducing a rotation-invariant learning scheme and extending it to a rotation-robust masked autoencoding framework that combines handcrafted rotation-invariant cues with self-supervised pretraining, enabling representations that are robust to arbitrary rotations while retaining useful pose information. Extensive experiments across multiple benchmarks show consistent robustness gains under noise and rotation without sacrificing clean-data performance, offering practical insights and solutions for deploying reliable point cloud models in real-world environments.