How Sampling Strategy Affects Imbalance Mitigation in LiDAR Segmentation: A Study of Structured vs. Random Point-Based Architectures

Imagine a world where self-driving cars and autonomous drones navigate by "seeing" with laser beams instead of eyes. These machines fire pulses of light that bounce off the ground, buildings, and trees, returning as a cloud of millions of individual dots. This is how they build a three-dimensional map of their surroundings. However, these maps are not perfectly balanced. In a typical city scene, the laser hits the vast, flat expanse of the road or the side of a building thousands of times more often than it hits a small, distant street sign or a pedestrian. This creates a severe imbalance: the computer sees the big things constantly but barely notices the small, critical details. If the machine's brain is not trained to handle this, it might confidently ignore a person standing on the sidewalk because it has been overwhelmed by the sheer volume of data about the road.

For decades, researchers have tried to fix similar problems in standard 2D photography, where some objects appear far less often than others. They developed various mathematical tricks to force the computer to pay more attention to the rare items. But when these same tricks were applied to the chaotic, three-dimensional world of laser point clouds, the results were confusing. Some methods worked, while others seemed to make the problem worse. The question remained: does the way a computer samples these laser dots change which tricks actually work? A new study by researchers at the University of Cyprus and the KIOS Research and Innovation Center of Excellence set out to answer this by testing a wide range of solutions on real-world data.

The researchers focused on two different ways computers process these laser clouds. The first method, used in a system called KPConv, acts like a careful surveyor. It picks points in a structured, organized pattern, ensuring that every part of the scene is sampled with a specific geometric logic. The second method, used in a system called RandLA-Net, is much more chaotic; it grabs points completely at random, like a child grabbing handfuls of sand from a beach. The team tested both systems against three very different datasets: a massive aerial scan of a city with an extreme imbalance of 641 to one, an indoor scan of a building with a moderate imbalance, and a synthetic, computer-generated environment. They applied eleven different strategies to see which one helped the computer recognize the rare objects best.

The results delivered a surprising and clear verdict. The most common instinct in the field—simply telling the computer to weigh the rare objects more heavily based on how often they appear—turned out to be a disaster. When the researchers used this standard "inverse-frequency" weighting, the computer's performance dropped significantly, sometimes by as much as 12 percent. In the worst cases, the system failed catastrophically, completely missing small objects like poles and fences. The study explicitly ruled out this popular approach, showing that blindly boosting the importance of rare classes actually confuses the model and degrades its ability to see the world correctly.

Instead, the study found that the best solution depends entirely on how the computer is sampling the data. For the structured, surveyor-like system, the simplest approach was often the best. Using a uniform weighting, where every class is treated equally without any special mathematical adjustments, performed almost as well as the most complex, custom-designed formulas. The difference was so small—less than 2 percent—that the extra effort to calculate complex weights offered no real benefit. The structured sampling method seemed to naturally handle the imbalance well enough that the computer did not need to be forced to pay attention to the rare items.

However, the story was different for the random, chaotic sampling method. Because this system discards the natural geometric structure of the scene, it is much more sensitive to the imbalance. Here, the simple uniform approach fell short, trailing behind specialized techniques by up to 4.6 percent. For this type of architecture, the researchers found that a specific loss function called LDAM, which adjusts how the computer learns from its mistakes, provided a noticeable improvement. Yet, even here, caution was required. Other complex formulas that worked well in 2D photography caused the random system to fail just as badly as the simple weighting did in the structured system.

The researchers also looked at the "shape" of the learning process to understand why these differences occurred. They found that for the structured system, the difficulty of the learning task was tied to how imbalanced the data was; the more unbalanced the data, the sharper and more difficult the learning path became. But for the random system, the difficulty was driven by the physical complexity of the scene itself. In cluttered indoor environments with many thin objects like chairs and tables, the learning path became jagged and hard to navigate, regardless of how many rare objects were present. This suggests that the random method struggles because it loses the structural context that helps the computer make sense of the scene.

Ultimately, this work provides a practical guide for engineers building these systems. It suggests that if you are using a structured sampling method, you should avoid the temptation to over-engineer your solution; a simple, equal treatment of all classes is robust and effective. If you are using a random sampling method, you do need a specialized tool to handle the imbalance, but you must choose it carefully, as the wrong tool can break the system entirely. The study concludes that there is no single magic bullet for all 3D vision tasks. Instead, the effectiveness of any solution is shaped by a delicate interaction between how the data is collected, how the computer samples it, and the specific nature of the environment it is trying to understand.