Revisiting a Simple MLP Framework for Z-Axis Rotation-Invariant Point Cloud Place Recognition

The objective of point cloud place recognition is to convert a point cloud into a global descriptor that can be utilized in autonomous driving applications to identify the best-matched road scene from an extensive dataset. However, capturing a point cloud from an arbitrary view by robots or self-driving vehicles often involves scene rotations, making existing deep learning-based methods susceptible to errors. To quantify this performance degradation, we introduce a novel metric: Average Recall@N under arbitrary rotations, denoted as “R-AR@N.” To address this issue, we propose a geometrical transformation module designed to convert rotation-sensitive coordinates into rotation-invariant representations. Additionally, we observe that the design of overly complex networks may not be crucial for effective point cloud analysis. In line with the straightforward architectural design of PointMLP [20], we introduce a local feature transformation module that utilizes statistical representations to transform local point features within a reasonable range. This enables the network to capture diverse geometric structures and generate a robust global descriptor. Our proposed method undergoes extensive evaluation on the Oxford outdoor dataset and three in-house datasets, demonstrating an improvement of at least 2% over previous methods on the newly proposed “R-AR@N” metric. Our code is available at https://github.com/jasonwjw/RI-PointMLP.