Overcoming Local Minima in Iterative Point Registration

577 concepts460 papers8 proposals

Imagine trying to align two identical 3D shapes. If they are already close to each other, a simple greedy strategy works beautifully: match each point on the first shape to its nearest neighbor on the second, compute the optimal transformation to minimize the distance between these pairs, and step closer. This intuitive loop is the foundation of the classic Iterative Closest Point (ICP) algorithm.

But what happens if one shape is rotated 180 degrees or shifted far away? The nearest-neighbor heuristic breaks down completely. Points on the left side of one object match with the right side of the other, generating deceptive geometric forces that pull the algorithm into a suboptimal, scrambled alignment.

This is the core challenge of initialization sensitivity in spatial registration. Traditional local registration techniques operate within a limited "capture range"—a narrow basin of attraction surrounding the true alignment. When the initial displacement is large, local optimizers inevitably get trapped in false local minima, failing to find the globally optimal match.

Local Optimization Landscape:
           False Minimum             True Global Minimum
            (Local Trap)             (Perfect Alignment)
               \     /                     \     /
                \   /                       \   /
  _______________\_/_________________________\_/_______________
  Initial Pose (Far Away) ----> Trapped!

The Chicken-and-Egg Trap of Correspondence

To understand why registration algorithms are so sensitive to their starting positions, Li frames 3D registration as a fundamental chicken-and-egg dilemma:

  1. To find the correct spatial transformation, you must first know which points correspond to each other.
  2. To find the correct point correspondences, you must first know the relative spatial transformation.

When an algorithm is initialized far from the ground truth, the initial correspondence estimates are highly inaccurate. As pointed out by Phillips et al., this problem is severely compounded by outliers and noise. Distant, unrelated points can exert a strong pull during the early stages of optimization, further shrinking the effective capture range and ensuring that local methods converge to incorrect local optima.

This brittleness makes pure local optimization impractical for autonomous systems, medical imaging, or large-scale mapping where a reliable starting guess is rarely guaranteed. To solve this, researchers have developed three primary strategies to bypass or flatten these local traps.

Instead of trying to navigate a treacherous optimization landscape from a single starting point, why not search the entire space systematically?

Algorithms like Go-ICP, developed by Yang et al., and GO-APSR, proposed by Li et al., employ a Branch-and-Bound (BnB) framework to guarantee global optimality. Instead of guessing a single trajectory, the BnB approach divides the 3D transformation space into geometric regions (cubes). For each region, the algorithm computes a mathematical lower bound on the alignment error. If a region's best possible error is worse than an alignment already found elsewhere, that entire branch of the search space is discarded ("pruned"). This allows the algorithm to find the absolute global minimum regardless of how poorly the shapes were initially positioned.

Alternatively, the problem can be reframed to avoid iterative search entirely. Ho et al. introduced an algebraic approach that matches the statistical moments of the underlying point distributions. By converting registration into a polynomial optimization problem, they leverage algebraic geometry to solve for the global minimum directly, rendering the algorithm completely immune to initial spatial orientation.

Strategy 2: Decoupling the Transformation Space

A rigid 3D transformation involves six degrees of freedom (3D rotation and 3D translation) interacting simultaneously, creating a highly non-convex search space. An intuitive way to widen the capture range is to break this complex space into simpler, independent components.

For example, Liu et al. developed a global translation search method (GoTS) that utilizes Rotation Invariant Features (RIF). Because these features remain identical regardless of how the object is rotated, the algorithm can completely ignore rotation in its first phase. This collapses the grueling 6D optimization landscape into a much simpler 3D translation search, drastically reducing the risk of getting stuck in local minima. Once the translation is locked in, estimating the remaining rotation becomes straightforward.

A parallel intuition is used in unsupervised deep learning. Zhang et al. designed a network that explicitly separates "pose-invariant" structural features from "pose-related" features. By isolating the global shape geometry from the current spatial orientation, the model prevents extreme initial misalignments from corrupting its internal representations, allowing it to establish a common reference frame from arbitrary starting poses.

Strategy 3: Smoothing the Optimization Landscape

Traditional ICP relies on hard, discrete point-to-point assignments. If a point matches the wrong neighbor, the objective function jumps discontinuously, creating a jagged optimization landscape riddled with micro-local minima. Modern approaches eliminate these traps by smoothing out the landscape itself.

Discrete Matching (Jagged, many local traps):
  /\_/\_/\_/\_/\_/\_/\

Continuous/Probabilistic Matching (Smooth, wide basin of attraction):
  \_________________/

One way to achieve this smoothness is to move from raw spatial coordinates to continuous feature spaces. Huang et al. introduced feature-metric registration, which optimizes alignment by minimizing intensity differences in a learned, continuous feature space. Because the features vary smoothly across space, the objective function loses its jagged local traps, creating a massive basin of attraction that can tolerate large initial offsets.

From a probabilistic perspective, algorithms can treat point clouds as continuous probability distributions rather than collections of hard coordinates. Gao et al. formulated this using Gaussian filtering (FilterReg), while Mi et al. utilized regularized Wasserstein distances. By computing soft, probabilistic correspondences across the entire distribution, these methods ensure that every point pair contributes a smooth, continuous gradient. A large initial rotation no longer causes a catastrophic misassignment; instead, the global statistical properties of the shapes gently guide the system toward the true global optimum.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 447 papers
  2. Version 1 · 2026-10-04 · 629 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_initialization_sensitivity_local_minima__3829488a,
  title = {initialization sensitivity (local minima sensitivity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/initialization-sensitivity-local-minima-sensitivity/rdr_3829488a34ff4b50b2a3ee688c6181d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_initialization_sensitivity_local_minima__3829488a_v2,
  title = {initialization sensitivity (local minima sensitivity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/initialization-sensitivity-local-minima-sensitivity/rdr_3829488a34ff4b50b2a3ee688c6181d4/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake