Assessing the practical applicability of neural-based point clouds registration algorithms: A comparative analysis

Point cloud registration is a vital task in three-dimensional (3D) perception, with several different applications in robotics. Recent advancements have introduced neural-based techniques that promise enhanced accuracy and robustness. In this paper, we thoroughly evaluate well-known neural-based point cloud registration methods using the Point Clouds Registration Benchmark, which was developed to cover a large variety of use cases. Our evaluation focuses on the performance of these techniques when applied to real-complex data, which presents a more challenging and realistic scenario than the simpler experiments typically conducted by the original authors. The results reveal considerable variability in performance across different techniques, highlighting the importance of assessing algorithms in realistic settings. Notably, 3DSmoothNet emerges as a standout solution, demonstrating good and consistent results across various data sets. Its efficacy, coupled with a relatively low graphics processing unit (GPU) memory footprint, makes it a promising choice for robotics applications, even if it is not yet suitable for real-time applications due to its execution time. Fully Convolutional Geometric Features also performs well, albeit with greater variability among data sets. PREDATOR and GeoTransformer are promising, but demand substantial GPU memory, when handling large point clouds from the Point Clouds Registration Benchmark. A notable finding concerns the performance of Fast Point Feature Histograms, which exhibit results comparable to the best approaches while demanding minimal computational resources. Overall, this comparative analysis provides valuable insights into the strengths and limitations of neural-based registration techniques, both in terms of the quality of the results and the computational resources required. This helps researchers to make informed decisions for robotics applications.