VNI-Net: Vector neurons-based rotation-invariant descriptor for LiDAR place recognition

Abstract

Despite the emergence of various LiDAR-based place recognition methods, the challenge of place recognition failure due to rotation remains critical. Existing studies have attempted to address this limitation through specific training strategies involving data augment and rotation-invariant networks. However, augmenting 3D rotations ($\mathrm{SO}\left(3\right)$) is impractical for the former, while the latter primarily focuses on the reduced problem of 2D rotation ($\mathrm{SO}\left(2\right)$) invariance. Existing methods targeting $\mathrm{SO}\left(3\right)$ rotation invariance suffer from limitations in discriminative capability. In this paper, we propose a novel approach (VNI-Net) based on the Vector Neurons Network (VNN) to achieve $\mathrm{SO}\left(3\right)$ rotation invariance. Our method begins by extracting rotation-equivariant features from neighboring points and projecting these low-dimensional features into a high-dimensional space using VNN. We then compute both Euclidean and cosine distances in the rotation-equivariant feature space to obtain rotation-invariant features. Finally, we aggregate these features using generalized-mean (GeM) pooling to generate the global descriptor. To mitigate the significant information loss associated with formulating rotation-invariant features, we propose computing distances between features at different layers within the Euclidean space neighborhood. This approach significantly enhances the discriminability of the descriptors while maintaining computational efficiency. We conduct experiments across multiple publicly available datasets captured with vehicle-mounted, drone-mounted LiDAR sensors and handheld. VNI-Net outperforms baseline methods by up to 15.3% in datasets with rotation, while achieving comparable results with state-of-the-art place recognition methods in datasets with less rotation. Our code is open-sourced at https://github.com/tiev-tongji/VNI-Net.