Centroid Anchoring for Stable Object Representation

18 concepts12 papers1 proposal

In computer vision, how we choose to anchor an object's coordinate system dictates how well a model can handle real-world messiness like occlusions, non-rigid movements, and irregular shapes. For years, the default strategy has been to use bounding boxes and treat the center of the box as the object's origin.

While bounding boxes are computationally convenient, their centers are inherently fragile. If an object becomes partially occluded, or if an articulated object (like a person bending over) changes its posture, the bounding box shifts drastically. As a result, the calculated "center" moves away from the actual physical core of the object.

To solve this, a powerful paradigm has emerged: centroid-based anchoring. Instead of relying on the arbitrary spatial boundaries of a bounding box, models locate and anchor themselves to the object's true geometric centroid or center of mass. This shift provides a fundamentally more stable reference point for shape reconstruction, pose estimation, and instance segmentation.

The Fragility of Bounding Box Centers

To understand why centroid anchoring works, we must first look at why bounding boxes fail. A bounding box is determined entirely by an object's extreme peripheral points. If a target is partially hidden behind an obstacle, its visible boundaries change, causing the bounding box center to jump unpredictably.

As Yang Zheng et al. demonstrate in DPPD: Deformable Polar Polygon Object Detection, this instability severely undermines shape reconstruction. When the coordinate origin shifts due to occlusion, any geometric predictions tied to that origin become inaccurate.

Furthermore, for highly irregular or non-convex shapes (such as an L-shaped object or a crescent moon), the geometric center of a bounding box can actually fall completely outside the object's physical boundaries. Cewu Lu et al. note in Contour Box: Rejecting Object Proposals without Explicit Closed Contours that when an object's true center aligns poorly with its bounding box center, optimization and contour detection frameworks struggle to accurately capture the intended boundaries.

Geometric Leverage and Positional Stability

Centroid anchoring solves this by placing the reference frame deep within the dense core of the object. This positioning leverages a basic principle of geometry: minimizing geometric leverage.

Peripheral Anchor (High Leverage Error)
[X]-------------------------------------> [Feature]
A small error in the anchor position [X] causes a massive angular shift.

Centroid Anchor (Low Leverage Error)
                 [X]-----> [Feature]
Centrally located anchors minimize the distance to all features, reducing error amplification.

In Detecting and Localizing 3D Object Classes using Viewpoint Invariant Reference Frames, Matthew Toews et al. establish that reference frames located centrally within an object's structure offer significantly higher geometric stability than peripheral ones. Because central points minimize the average distance to all other features, localized measurement errors have less "leverage" to cause massive shifts in the overall predicted location.

This property is equally valuable for non-rigid or articulated shapes. When a 3D object moves or bends, its surface vertices fluctuate wildly. Fiora Pirri et al. address this in Consistent pose normalization of non-rigid shapes using One-Class Support Vector Machines by calculating the centroid of the volume contained within a learned decision boundary. By anchoring the coordinate system to this dense volumetric core rather than surface vertices, the authors significantly reduce the positional jitter caused by moving limbs.

This principle also extends to industrial applications. For example, Jianxin Feng et al. show in Laser weld spot detection based on MMG-YOLO that calculating the center of mass rather than a box center provides stable localization for objects with highly irregular morphological features.

Polar Coordinate Modeling and the Inside-the-Shape Property

The stability of centroid anchoring is perhaps most vividly realized in polar-coordinate-based object detection and segmentation. Instead of predicting pixel masks or bounding box corners, models like PolarMask parameterize an object by identifying a central origin and shooting out radial rays at fixed angular intervals to map the contour.

      Polar Ray Regression
             .  :  .
           .    |    .
          .  --(*)--  .  <-- Origin (*) placed at Center of Mass
           .    |    .
             .  :  .

For this mathematical formulation to work reliably, the origin must reside inside the object's boundary. If the origin falls outside the shape, a single angular ray could cross the boundary multiple times or miss it entirely, making radial length estimation an ill-posed problem.

To guarantee that the origin remains safely inside the object, Enze Xie et al. design PolarMask: Single Shot Instance Segmentation With Polar Representation to explicitly use the object's center of mass as the coordinate origin. This design choice ensures the reference point remains within the object's boundaries, stabilizing radial regression for complex or non-convex shapes.

The authors further refine this concept in PolarMask++: Enhanced Polar Representation for Single-Shot Instance Segmentation and Beyond by introducing a "centerness" module. This module weights candidate points based on their proximity to the true mass center, ensuring that only the most geometrically stable points are used to generate contour predictions.

Similarly, in the context of mathematical shape analysis, Rui Li et al. demonstrate in A new area-based convexity measure with distance weighted area integration for planar shapes that using a stable central reference point—such as the geometric center of a shape's convex hull—allows for robust internal area integration that remains resilient against digital noise.

Overcoming Occlusion and Fragmentation

Centroid anchoring also fundamentally changes how models handle multi-object grouping and fragmentation. In crowded environments, objects frequently cut in front of one another, visually splitting a single instance into disconnected pieces.

Traditional bounding-box methods struggle here because overlapping boxes lead to ambiguous detections, often requiring complex post-processing like Non-Maximum Suppression (NMS) to clean up duplicates. However, NMS frequently fails when objects naturally overlap closely. As Rui Yang et al. show in UniInst: Unique Representation for End-to-End Instance Segmentation, identifying unique, instance-aware points instead of spatial boxes allows models to cleanly differentiate between distinct objects even when they occupy nearly identical regions of an image.

When an object is fragmented by an obstacle, centroid anchoring acts as a global coordinate glue. In Occlusion-Resistant Instance Segmentation of Piglets in Farrowing Pens Using Center Clustering Network, Endai Huang et al. demonstrate that predicting central coordinates allows a system to maintain a single identity for an animal, even when its body is physically bisected by a pen bar. Because the model targets the underlying center, it avoids generating duplicate detections for the separated visible parts.

  Fragmented Object Grouping
   [ Visible Part A ]   ===\
        (Obstacle)          --->  Regress to Shared Centroid [X]
   [ Visible Part B ]   ===/

This capability is vital for human pose estimation, where limbs frequently cross or disappear behind objects. In Pose Partition Networks for Multi-person Pose Estimation, Xuecheng Nie et al. utilize a person's center as a global anchor. Different visible body parts can independently regress vectors back to this shared centroid, ensuring that fragmented joint detections are correctly clustered into the correct individual.

The authors expand on this in Single-Stage Multi-Person Pose Machines, proving that using the instance centroid as a root anchor provides powerful resistance against extreme deformations. Because the centroid acts as a stable, coordinate-independent origin, it simplifies the complex task of associating scattered anatomical landmarks with their respective individuals in a single forward pass.

Key Foundational Work

Detecting and Localizing 3D Object Classes using Viewpoint Invariant Reference Frames (2007) — contributed 1 concept elements to this direction.
Consistent pose normalization of non-rigid shapes using One-Class Support Vector Machines (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_centroid_based_anchoring_58adaead,
  title = {centroid-based anchoring},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/centroid-based-anchoring/rdr_58adaeade9e9484a972e44e62f2cf885},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_centroid_based_anchoring_58adaead_v1,
  title = {centroid-based anchoring},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/centroid-based-anchoring/rdr_58adaeade9e9484a972e44e62f2cf885/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake