SCIRCNet: Rotation-Invariant Point Cloud Completion via Self-Contour-Aware Representation

Abstract

In real-world scenarios, point clouds can take on diverse poses and suffer from incompleteness. Point cloud completion benefits from rotation-invariant feature representations, which remain largely unaddressed in current study. To this end, we propose a Self-Contour Invariant Representation-based Completion Network (SCIRCNet), a novel framework designed to robustly complete partial point clouds regardless of arbitrary 3D rotation. First, an SCIR module is introduced to transform point clouds into a rotation-invariant representation based on intrinsic geometric extremes. This module effectively eliminates the dependency on absolute coordinate systems, creating a stable foundation for subsequent feature learning. Second, a dual-branch feature extraction strategy is designed to integrate a Class Discriminant Feature Extractor (CDFE) and a Geometry-Aware Feature Extractor (GAFE). The CDFE aligns features into a unified intra-class framework to capture category-specific semantics, while the GAFE utilizes adaptive neighborhoods to extract fine-grained geometric details. This dual-branch strategy synergizes global shape coherence with local geometric fidelity. Finally, a three-stage coarse-to-fine pipeline is developed to progressively recover high-fidelity 3D shapes. By initially establishing global structural priors to synthesize a coarse completion, and subsequently refining local details via a folding-based mechanism, this pipeline ensures structurally complete and visually plausible outputs. Extensive experiments on three benchmark datasets demonstrate that SCIRCNet achieves superior performance compared to existing approaches, particularly demonstrating exceptional robustness against arbitrary 3D rotation. Note to Practitioners—This paper addresses the challenge of completing partial point cloud data captured with arbitrary orientations in real-world scenarios. In practical workflows, such as robotic bin picking, the requirement for objects to be perfectly aligned to a canonical coordinate system is inevitably computationally prohibitive. Standard reconstruction algorithms typically fail when input data does not match a pre-defined orientation. To address this, this paper proposes a geometry-aware method that eliminates the dependency on external coordinate alignment. The method establishes a reference frame derived from the object’s intrinsic geometric extremes rather than absolute coordinates, enabling robust shape completion regardless of rotation changes. For practitioners, this method streamlines the deployment pipeline by eliminating the dependency on complex pose estimation modules. It is particularly valuable for automated industrial inspection, robotic manipulation in unstructured environments, and autonomous vehicle perception, where input data is inherently sparse and unaligned.