SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration

Abstract: Extracting robust and general 3D local features is key to downstream tasks such as point cloud registration and reconstruction. Existing learning-based local descriptors are either sensitive to rotation transformations, or rely on classical handcrafted features which are neither general nor representative. In this paper, we introduce a new, yet conceptually simple, neural architecture, termed SpinNet, to extract local features which are rotationally invariant whilst sufficiently informative to enable accurate registration. A Spatial Point Transformer is first introduced to map the input local surface into a carefully designed cylindrical space, enabling end-to-end optimization with SO(2) equivariant representation. A Neural Feature Extractor which leverages the powerful point-based and 3D cylindrical convolutional neural layers is then utilized to derive a compact and representative descriptor for matching. Extensive experiments on both indoor and outdoor datasets demonstrate that SpinNet outperforms existing state-of-the-art techniques by a large margin. More critically, it has the best generalization ability across unseen scenarios with different sensor modalities. The code is available at https://github.com/QingyongHu/SpinNet.

• The paper introduces SpinNet, a novel method using a spatial point transformer and neural feature extractor to create rotation-invariant descriptors for 3D point cloud registration.
• The paper achieves an average recall of 92.8% on the ETH dataset, outperforming previous methods by nearly 13% even when trained solely on indoor data.
• The paper demonstrates that robust descriptor design can enhance 3D computer vision applications such as robotics and autonomous driving.

Overview of "SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration"

The paper "SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration" introduces a novel approach to extracting local features from 3D point clouds that are both rotationally invariant and informative enough for accurate registration tasks. The ability to robustly match parts of 3D surfaces is crucial in applications such as point cloud registration, reconstruction, and segmentation.

Core Contributions

The authors present a method called SpinNet, which effectively overcomes the limitations faced by previous learning-based local descriptors. These limitations include sensitivity to rotational transformations and reliance on classical handcrafted features that are generally not representative enough for diverse scenarios.

The primary innovations in SpinNet consist of:

1. Spatial Point Transformer: This component maps the input local surface into a specially designed cylindrical space. This design leverages an SO(2) equivariant representation, allowing for point clouds to be transformed in a way that maintains their rotational invariance along the radial axis, while supporting end-to-end optimization.
2. Neural Feature Extractor: By employing point-based and 3D cylindrical convolutional neural layers, this module extracts a compact and representative descriptor for feature matching. The approach ensures that the learned descriptors are both detailed and generalizable, optimizing the learning of local patterns without relying on handcrafted features.

Experimental Results

The SpinNet demonstrates significant improvements over state-of-the-art techniques in various datasets. For example, it shows an average recall score of 92.8% in the ETH dataset, outperforming the previous best by almost 13%, despite being trained only on indoor 3DMatch data. This result highlights SpinNet's ability to generalize across different environments and sensor modalities, marking a significant step forward in 3D registration tasks.

The analyses presented compare SpinNet's descriptors with both rotation-variant and rotation-invariant approaches. The performance gain is evident in its capability to consistently handle rotational changes and noise typical of real-world 3D point clouds.

Implications and Future Directions

Theoretical implications of this research reveal important insights into designing neural architectures that address rotational variance without compromising the richness of feature representation. Practically, SpinNet shows promise in enhancing 3D computer vision tasks that demand robust registration and matching capabilities across a range of scenarios and data conditions, such as autonomous driving and robotics.

Possible future directions include the integration of the descriptor with advanced keypoint detection strategies and adaptation to fully-convolutional architectures, which could further enhance the robustness and efficiency of the registration process.

In conclusion, SpinNet provides a compelling solution to 3D point cloud registration challenges, offering a robust, generalizable approach through innovative spatial transformations and neural feature extraction. It sets a new benchmark for performance in both familiar and unseen environments, showing versatility and advanced understanding of surface descriptors in 3D space.