MemReg Outperforms Pairwise Methods and SLAM in LiDAR Point Cloud Registration

A team from the IRMVLab lab has proposed MemReg, a neural network for LiDAR point cloud registration that treats a scan not as one half of an isolated pair of frames but as an element of a continuous sequence. It is built on two hybrid memory buffers and a history-weighting module. The method was tested on three outdoor datasets: KITTI, nuScenes and Apollo-Southbay. According to the authors, it outperforms all pairwise methods and multi-frame SLAM systems. The preprint is available on arXiv, and the authors promise to publish the code on GitHub.

The problem with pairwise registration

Point cloud registration determines the relative position and orientation of two scans. Odometry, mapping and localization for self-driving cars and robots are built on it. Most neural network solutions for outdoor scenes take two frames and output the transformation between them. They ignore everything that came before those frames.

Where the data is good, this works. But a real LiDAR produces sparse measurements, and these are also corrupted by outliers: passing cars, pedestrians, reflections. When neighboring scans overlap only slightly, for example during a sharp turn or on an empty highway, a pairwise model has nothing to anchor to. The MemReg authors start from a simple observation: a long temporal sequence contains global context. It can be used to fill in sparse measurements, filter out outliers and cope with low overlap.

How the memory works

The architecture relies on two buffers of different types.

• The implicit feature buffer stores pose-related latent representations that the network extracted from previous frames.
• The explicit pose buffer stores poses that have already been estimated, that is, concrete geometric information.

Both buffers update dynamically as new scans arrive. This is closer to streaming than to batch processing: the model does not recompute the entire history but maintains a compact state.

Another component is the dynamic history-weighting module. It adaptively decides how heavily the current frame should rely on accumulated features. The task is nontrivial: old information helps with noise and low overlap, but it can hurt if the scene changes abruptly or if error has already built up. The module manages this balance without hard-coded rules.

What the experiments showed

The authors report the following results:

• on KITTI, nuScenes and Apollo-Southbay, MemReg reaches state-of-the-art level;
• the method outperforms all previous pairwise approaches;
• it also beats multi-frame SLAM systems, which likewise use history but are designed differently;
• indoors, the method delivers competitive results on 3DMatch, 3DLoMatch and ScanNet, even though it was designed for outdoor use.

The abstract gives no specific metric values, so the size of the margin can only be assessed once the full paper and code are released. This is an important caveat. The claim that it "outperforms all" methods needs to be checked by independent reproductions, especially for runtime and memory consumption, which the authors do not mention in the abstract.

Practical takeaways

Teams building autonomous driving, warehouse robots or mapping platforms should note two things. First, the gain comes from the problem formulation itself: moving from pairs to a stream adds context without changing the sensor. Second, the indoor results hint that the approach transfers across domains, though with caveats.

Before adopting it, wait for the IRMVLab repository release and check three parameters: latency on the target hardware, buffer size on long routes, and behavior when several frames in a row are lost. These are the points that separate a good research architecture from a module that can be put into a real navigation stack.