Abstract
Point cloud recognition is an essential task in industrial robotics and autonomous driving. Recently, several point cloud processing models have achieved state-of-the-art performances. However, these methods lack rotation robustness, and their performances degrade severely under random rotations, failing to extend to real-world scenarios with varying orientations. To this end, we propose a method named Self Contour-based Transformation (SCT), which can be flexibly integrated into various existing point cloud recognition models against arbitrary rotations. SCT provides efficient rotation and translation invariance by introducing Contour-Aware Transformation (CAT), which linearly transforms Cartesian coordinates of points to translation and rotation-invariant representations. We prove that CAT is a rotation and translation-invariant transformation based on the theoretical analysis. Furthermore, the Frame Alignment module is proposed to enhance discriminative feature extraction by capturing contours and transforming self contour-based frames into intra-class frames. Extensive experimental results show that SCT outperforms the state-of-the-art approaches under arbitrary rotations in effectiveness and efficiency on synthetic and real-world benchmarks. Furthermore, the robustness and generality evaluations indicate that SCT is robust and is applicable to various point cloud processing models, which highlights the superiority of SCT in industrial applications.
Introduction
While deep learning has revolutionized 3D point cloud recognition tasks like classification and segmentation, most current models suffer from a critical lack of rotation robustness. Standard architectures, such as PointNet or DGCNN, are typically trained on "aligned" datasets where objects share a common orientation. In real-world scenarios like autonomous driving or industrial robotics, sensors capture objects at arbitrary orientations. Without extensive and computationally expensive data augmentation, the performance of these models collapses when faced with random rotations.
Existing rotation-invariant methods often transform coordinates into relative distances or angles. However, these non-linear transformations can lose global geometric structure or create "shape ambiguity," where distinct objects appear identical to the network. There is a clear need for a transformation that is both invariant and geometry-preserving.
Methodology
The proposed Self Contour-based Transformation (SCT) framework is a "plug-and-play" module that can be integrated into various existing point cloud recognition models. It consists of two primary components:
- Contour-Aware Transformation (CAT): Unlike non-linear methods, CAT uses a linear transformation to define a new coordinate system based on the object's own geometry. It calculates the object's barycenter and identifies the farthest and closest points to establish an orthogonal basis. Because this basis moves with the object, the resulting representation is strictly invariant to the original orientation while preserving global geometric structure.
- Frame Alignment (FA) Module: While CAT ensures invariance, it may leave objects in poses that are difficult for a network to learn. The FA module uses a small encoder-decoder structure to capture contour features and regresses a quaternion to rotate the object into a consistent "intra-class" frame (e.g., ensuring all chairs are oriented upright relative to each other), facilitating easier discriminative feature extraction.
Experimental Results
The researchers evaluated SCT across major benchmarks including ModelNet40 (synthetic), ScanObjectNN (real-world scans), and ShapeNet (part segmentation).
- Accuracy Recovery: On ModelNet40, a standard DGCNN model's accuracy drops from 92.9% to 20.6% under random rotations. Integrating SCT brings this accuracy back to 89.5% without any extra rotation-based training.
- Real-World Robustness: On the challenging ScanObjectNN dataset, which contains noise and occlusions, SCT outperformed all other rotation-invariant competitors, achieving 84.6% accuracy under random rotations.
- Efficiency: SCT maintains a lower computational complexity ($O(N)$) compared to other invariant methods that rely on expensive K-nearest neighbor searches ($O(NK)$).
- Segmentation: In part segmentation tasks on ShapeNet, SCT maintained an 81.4% mIoU regardless of rotation, whereas standard models saw their performance cut in half.
Conclusion
SCT provides a robust, efficient, and portable solution for rotation-invariant point cloud recognition. By leveraging linear transformations and learned frame alignment, it preserves geometric integrity while achieving state-of-the-art performance. Its "plug-and-play" nature makes it highly applicable to various industrial and robotic sensing tasks where orientation cannot be guaranteed. Future research may extend these self-contour-based frames to large-scale 3D object detection in outdoor scenes.
Summary updated