Geometric Centroid Voting in 3D Point Clouds

This direction is split into other directions: point-to-center offset estimation, geometric feature dominance. Its last version stays readable and citable here.

Version 1 of 163 papers79 conceptsfrozen 2026-10-04

In 3D computer vision, identifying individual objects within a point cloud is notoriously difficult because sensors only capture the "skin" of the world. Whether using LiDAR, Radar, or RGB-D cameras, we receive a sparse shell of points on the surfaces of objects, leaving the interiors empty.

This surface-only data creates two major hurdles for traditional clustering. First, points on opposite sides of a large object (like a car) are spatially far apart, making it hard to realize they belong together. Second, points on two different objects that are touching (like a person leaning against a wall) are spatially close, making them nearly impossible to separate.

To solve this, researchers have converged on a powerful geometric intuition: don't cluster the points where they are; cluster them where they belong. By training models to predict the geometric center of the object each point belongs to, the field has moved toward a "voting" architecture that collapses sparse surfaces into dense, easily separable centroids.

The Shift to Offset Space

The core mechanism, often called Center Shift Vectors (CSV) or Radial Voting, involves a dual-prediction task. For every point in a scene, the neural network predicts both a semantic label (e.g., "chair") and a 3D offset vector. This vector represents the displacement required to move that point from the object's surface to its estimated geometric center.

As Li Jiang and colleagues demonstrated with PointGroup, this transformation effectively "shrinks" object representations. When these offsets are applied, points that were once scattered across a wide surface area migrate toward a common coordinate. This creates artificial void space between adjacent objects of the same class, allowing simple density-based clustering to succeed where it previously failed.

This "point-collapse" mechanism, as described by George Retsinas in the context of mushroom segmentation, simplifies instance segmentation by converting a high-dimensional embedding problem into a straightforward spatial clustering task.

Why Voting Beats Direct Regression

In cluttered scenes with heavy occlusion, we might only see a small fraction of an object—perhaps just the handle of a mug or the corner of a car. Direct regression (trying to guess the object's bounding box from that small fragment) is highly unstable.

Voting-based methods, such as those proposed by Yangzheng Wu and Michael Greenspan in RCVPose, treat every visible point as a "voter." Each point casts a vote for where it thinks the object's center or keypoints are. By aggregating these votes in a 3D accumulator grid, the model can identify "peaks" where the most predictions converge.

This approach is inherently robust to noise and occlusion. Even if half the points are missing or their individual predictions are slightly off, the consensus of the remaining points still points to the correct center. This geometric anchoring allows models to "hallucinate" the center of an object even when the center itself is empty space.

Geometry Over Appearance

A recurring theme across this research is that structure is more informative than style. While 2D vision relies heavily on color and texture, 3D models often perform better when they ignore visual appearance entirely.

Hema Swetha Koppula and Ashutosh Saxena found early on that geometric features like surface planarity and height from the floor are more effective for identifying indoor objects than color. More recently, Albert Mosella-Montoro showed that purely geometric spatial signatures can outperform multi-modal approaches that include RGB data.

This reliance on geometry is particularly critical in sparse domains like automotive radar. Jianan Liu and Weiyi Xiong demonstrated that predicting center shift vectors allows for high-accuracy instance segmentation even on extremely sparse radar points, where visual texture is non-existent.

Beyond Objects: Parts and Affordances

The intuition of voting for centers isn't limited to whole objects; it extends to functional parts and "affordances"—the parts of an object a robot can actually interact with.

In ArticuBot, Yufei Wang and colleagues use 3D displacement vectors to predict sub-goals for manipulating articulated objects like cabinets and drawers. Similarly, Xiaotong Chen uses Affordance Coordinate Frames to localize 3D keypoints for specific parts, like a spoon's scoop.

By shifting the focus from "what is this point" to "where is the center of the part this point belongs to," these models gain a physical understanding of the scene that is robust to changes in viewpoint, scale, and clutter. Whether the goal is to drive a car or stir a pot, the solution remains the same: use the surface to find the center.

Papers in this version

Showing 60 of 63 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geometric_centroid_voting_in_3d_point_cl_8b635055,
  title = {Geometric Centroid Voting in 3D Point Clouds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-centroid-voting-in-3d-point-clouds/rdr_8b6350559de44757ab4373b7fb79895c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_geometric_centroid_voting_in_3d_point_cl_8b635055_v1,
  title = {Geometric Centroid Voting in 3D Point Clouds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-centroid-voting-in-3d-point-clouds/rdr_8b6350559de44757ab4373b7fb79895c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}