Self-built dataset for better generalization in point cloud registration

Abstract

Abstract The training and testing datasets for point cloud registration networks often fail to satisfy the independent and identically distributed assumption, that actually leads to significant performance degradation and poor generalization. To address this problem, some point cloud data post-processing methods were proposed to construct a dataset with abundant domain discrepancy variables. Specifically, these variables were introduced into the training data by optimizing spatial sampling and temporal interval frame matching. A total of 63461 sets of point cloud registration data pairs were produced. A 31.8% improvement in accuracy was demonstrated compared to that of the benchmark dataset. The network generalization, quantized by the one-step generalization ratio, reached 0.9832, significantly improves the generalization and provides reference data for cross-domain research on point cloud registration.