Point cloud registration is a vital task in 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. With few exceptions, the results reveal a
significant performance gap, with most neural-based methods performing
poorly on real-complex data. However, amidst the underwhelming results,
3DSmoothNet stands out as an exception, demonstrating excellent
performance across the majority of benchmark sequences. Remarkably, it
outperforms a state-of-the-art feature extractor, namely FPFH,
showcasing its effectiveness in capturing informative and discriminative
features from point clouds. Given these results, we assert that while
neural-based techniques could have advantages over conventional methods,
further research is necessary to fully unlock their practical utility.
We advocate for more extensive studies employing unbiased and realistic
testing procedures, akin to the rigorous evaluation framework employed
in this work. This paper represents a crucial step towards establishing
a more robust literature in the field of point cloud registration. By
exposing the limitations and advantages of existing neural-based
methods, we aim to inspire future research and drive the development of
more effective techniques for real-world applications.