K-means Cluster Meshes: Mesh-Specific Algorithms

• K-means Cluster Meshes is a class of algorithms that adapt Lloyd’s method to mesh data by incorporating non-rigid transformations, geodesic distances, and graph centrality.
• The approaches replace the standard Euclidean centroid with mesh-specific representatives, such as deformation-invariant means or PageRank-based nodes, to better capture intrinsic data structures.
• Applications include mesh segmentation, balanced partitioning for parallel simulations, and enhanced clustering performance through soft balancing and efficient iterative updates.

K-means clustering on meshes comprises a set of adaptations of Lloyd’s centroid-based procedure to geometric data that are not adequately captured by a plain Euclidean mean. In the sources considered here, three mesh-specific settings are explicit: whole meshes represented by stacked vertex coordinate vectors $x_i \in \mathbb{R}^{3V}$; triangulated meshes treated as graphs $G=(V,E)$ with shortest-path or geodesic distance $d_G$; and large sets of mesh vertices $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$ partitioned for parallel numerical simulations. The corresponding methods preserve the intrinsic data space and augment K-means with a similarity measure invariant to non-rigid transformations, replace the Euclidean centroid with a graph-centrality–based representative, or adapt Lloyd’s heuristic with a soft balance mechanism (Cosentino et al., 2022, Hajij et al., 2020, Looz et al., 2018).

1. Problem formulations on mesh data

For clustering entire meshes with fixed vertex correspondence, the data are written as $X=\{x_1,\ldots,x_N\}$ with $x_i \in \mathbb{R}^{3V}$, where the $3D$ coordinates of $V$ vertices are stacked into one long vector. Spatial Transformer K-Means seeks $K$ centroids $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$ together with per-datum deformation parameters $G=(V,E)$0 and minimizes

$G=(V,E)$1

with

$G=(V,E)$2

Here $G=(V,E)$3 is a differentiable $G=(V,E)$4 spatial-transformer mapping, $G=(V,E)$5 is an optional regularizer on the deformation, and $G=(V,E)$6 trades off reconstruction against smoothness (Cosentino et al., 2022).

For mesh segmentation on a triangulation, the mesh is treated as an undirected graph whose vertices and edges come from the triangulation. On a general graph there is no vector mean, so the update step must choose a representative from the discrete set of vertices. In the PageRank-based formulation, one fixes $G=(V,E)$7, maintains seeds $G=(V,E)$8, forms the graph-Voronoi partition

$G=(V,E)$9

and updates each seed by centrality on the induced subgraph $d_G$0 (Hajij et al., 2020).

For geometric partitioning, the input is a set of mesh vertices, optionally with weights $d_G$1, and the objective combines the usual $d_G$2-means distortion with an explicit balance condition. The stated problem is to find a partition $d_G$3 of $d_G$4 that minimizes

$d_G$5

while enforcing

$d_G$6

for all $d_G$7, or, in the weighted case,

$d_G$8

Mesh partitioning is described as an indispensable tool for efficient parallel numerical simulations, with the goal of minimizing communication between the processes of a simulation while achieving load balance (Looz et al., 2018).

2. Lloyd-style iterations and mesh-specific representatives

All three approaches retain the alternating structure of Lloyd’s method, but they modify either the assignment rule, the representative, or both. Standard $d_G$9-means on a point set $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$0 alternates between a Voronoi-type assignment and an arithmetic-mean update. The graph formulation keeps the assignment idea but replaces the arithmetic centroid by a node-centrality–based representative, most prominently PageRank. After forming $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$1, one computes a PageRank vector on the induced subgraph and sets

$P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$2

The loop stops when there is no change in seeds or after a fixed iteration cap (Hajij et al., 2020).

Spatial Transformer K-Means follows a generalized Lloyd’s two-step scheme. In the E-step, for each $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$3 one solves

$P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$4

by gradient descent, computes

$P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$5

and assigns $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$6 to the cluster with minimal $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$7. In the M-step, for each cluster $P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$8 one updates

$P=\{x_1,\ldots,x_N\}\subset\mathbb{R}^d$9

which is stated to be the Fréchet mean in the warped space (Cosentino et al., 2022).

Balanced $X=\{x_1,\ldots,x_N\}$0-means extends the assign/move-centers scheme by introducing a per-cluster influence factor $X=\{x_1,\ldots,x_N\}$1 that penalizes oversized clusters. Instead of assigning a point $X=\{x_1,\ldots,x_N\}$2 to $X=\{x_1,\ldots,x_N\}$3, it assigns to the cluster minimizing

$X=\{x_1,\ldots,x_N\}$4

After each assignment sweep, cluster weights $X=\{x_1,\ldots,x_N\}$5 are compared to the target $X=\{x_1,\ldots,x_N\}$6, and the influence is adjusted as

$X=\{x_1,\ldots,x_N\}$7

with each adjustment capped to $X=\{x_1,\ldots,x_N\}$8 to avoid oscillation. Hamerly’s bounds are then used to skip distance computations when an assignment cannot change (Looz et al., 2018).

3. Spatial Transformer K-Means for non-rigid mesh clustering

Spatial Transformer K-Means is presented as a centroid-based clustering framework that jointly learns per-sample nonrigid alignments and cluster centroids in the original data space. The stated motivation is to preserve the intrinsic data space and augment K-means with a similarity measure invariant to non-rigid transformations. The method is claimed to enable the reduction of intrinsic nuisances associated with the data, reducing the complexity of the clustering task and increasing performances and producing state-of-the-art results; clustering in the input space of the data, leading to a fully interpretable clustering algorithm; and the benefit of convergence guarantees (Cosentino et al., 2022).

For mesh data, the non-rigid transform is a differentiable $X=\{x_1,\ldots,x_N\}$9 spatial-transformer mapping implemented by a thin-plate-spline or radial-basis-function warp driven by $x_i \in \mathbb{R}^{3V}$0 landmarks. In the thin-plate-spline construction, one chooses control points $x_i \in \mathbb{R}^{3V}$1 on a template or uniform grid in $x_i \in \mathbb{R}^{3V}$2. The warp $x_i \in \mathbb{R}^{3V}$3 is determined by solving the TPS linear system with kernel $x_i \in \mathbb{R}^{3V}$4 for the $x_i \in \mathbb{R}^{3V}$5 coefficients. The deformation vector $x_i \in \mathbb{R}^{3V}$6 consists of the offsets of the $x_i \in \mathbb{R}^{3V}$7 control points, with $x_i \in \mathbb{R}^{3V}$8, and $x_i \in \mathbb{R}^{3V}$9 is evaluated by applying $3D$0 to every vertex via barycentric or bilinear-like interpolation in $3D$1 (Cosentino et al., 2022).

In the simplest unregularized form $3D$2, the deformation-invariant distance is described as a quasipseudosemimetric: it is nonnegative, satisfies $3D$3 iff $3D$4 lies on the diffeomorphic-orbit of $3D$5, is invariant under the group of $3D$6 TPS-like diffeomorphisms, and is symmetric up to volume-change factors; a symmetrized version is possible if desired. Under the assumption that each inner problem has a unique global minimizer, the overall objective $3D$7 decreases at every iteration of Lloyd’s scheme and terminates at a local optimum in a finite number of steps. The stated justification invokes Banerjee et al. (2005)’s generalization of Bregman-divergence clustering, with $3D$8 treated as a valid Bregman-like divergence in the joint $3D$9 space (Cosentino et al., 2022).

The implementation recipe is explicitly mesh-oriented. Each mesh is represented by its stacked vertex coordinate vector in $V$0; connectivity remains fixed across the dataset, or one works on point-clouds when connectivity varies. Recommended preprocessing is to align all meshes to a common pose, for example with PCA-based orientation, normalize scale to unit volume, and optionally downsample high-resolution meshes to $V$1–$V$2 vertices. The per-$V$3 warp fit solves an $V$4 TPS system, plus $V$5 warp evaluation and $V$6 squared-error computation, leading to total per Lloyd iteration

$V$7

The practical ranges stated are $V$8–$V$9 and $K$0–$K$1, and GPU-parallelization across $K$2 can render the runtime nearly linear in $K$3 (Cosentino et al., 2022).

4. PageRank-based K-means on triangulated meshes

The PageRank formulation generalizes the $K$4-means update step from an arithmetic mean to a central node of a graph-defined cluster. For each Voronoi cell $K$5, one induces a subgraph $K$6 and computes a PageRank vector

$K$7

For undirected graphs, the standard definition recorded in the source is

$K$8

with damping factor $K$9, where $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$0 is the degree in $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$1. The new seed is then

$\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$2

Although no explicit single energy function is written down, the algorithm is described as alternately minimizing within-cluster distance in the assignment step and maximizing cluster-center centrality in the update step (Hajij et al., 2020).

The mesh-specific interpretation is direct: triangulated meshes enter as the graph $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$3 whose vertices are mesh vertices and edges are the mesh edges, possibly weighted by edge length, and $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$4 is the geodesic distance along the mesh. A naive Voronoi construction computes shortest-path distances from each seed to all vertices in each iteration; if $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$5 and $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$6, the stated cost is $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$7 via Dijkstra. For very large meshes, the “heat method” can be used in $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$8 per seed to approximate geodesics, and multi-source BFS is available when edges are unweighted. PageRank on each induced subgraph can be computed by power iteration in $\mu_1,\ldots,\mu_K \in \mathbb{R}^{3V}$9, so that for $G=(V,E)$00 and sparse meshes, each iteration is roughly $G=(V,E)$01 (Hajij et al., 2020).

The empirical results reported for meshes are qualitative rather than metric-based. The examples cited are a human head mesh segmented into $G=(V,E)$02 regions and a mechanical part segmented into $G=(V,E)$03 logical components. The resulting clusters are described as visually coherent and as respecting mesh geometry more robustly than naïve $G=(V,E)$04-means on vertex positions. No quantitative “purity” or “IoU” numbers are provided, but convergence is reported to occur in under $G=(V,E)$05 iterations, and geodesic-based Voronoi plus PageRank update is said to yield sharp, stable boundaries at concavities. The update to a topologically central vertex avoids the “mean–falling-off” problem in which an arithmetic centroid lies outside a cluster’s graph-connected component (Hajij et al., 2020).

5. Balanced K-means for mesh partitioning at scale

Balanced $G=(V,E)$06-means is introduced in the context of mesh partitioning for parallel numerical simulations. The stated comparison is that established graph-based partitioning tools yield a high solution quality, but their scalability is limited, whereas geometric approaches usually scale better, but their solution quality may be unsatisfactory for non-trivial mesh topologies. The method is the core of the partitioning algorithm Geographer, and bootstrapping of initial centers is performed with space-filling curves, leading to fast convergence of the subsequent balanced $G=(V,E)$07-means algorithm (Looz et al., 2018).

The key modification is the soft balancing mechanism based on the influence factor $G=(V,E)$08. Oversized clusters are penalized during assignment, and undersized clusters are relatively favored. The assign-and-balance routine initializes upper and lower bounds, computes local cluster sizes, performs an all-reduce to obtain global sizes, tests whether

$G=(V,E)$09

and otherwise updates each influence by

$G=(V,E)$10

with clamping to $G=(V,E)$11. The outer loop recomputes new centers by a global weighted average and tests convergence of center motion (Looz et al., 2018).

The implementation is explicitly parallel MPI. Each rank holds a contiguous block of points $G=(V,E)$12; centers and $G=(V,E)$13 are globally replicated on each rank; assignment is fully data-parallel on local points; a single all-reduce over $G=(V,E)$14 values communicates block weights after each assign-and-balance sweep; and another all-reduce communicates per-cluster coordinate sums to form the new $G=(V,E)$15. Running time per iteration is given as roughly

$G=(V,E)$16

plus the cost of two global reductions of length $G=(V,E)$17 (Looz et al., 2018).

The reported empirical range is extensive: 2D benchmark meshes from the 10th DIMACS challenge, 2.5D climate meshes, 3D ALEA testcases, and Delaunay triangulations from $G=(V,E)$18 to $G=(V,E)$19 vertices. The metrics are edge cut, max and total communication volume, block diameter, measured SpMV communication time, and total partitioning time. The aggregate findings state that on 2D DIMACS meshes, Geographer’s total communication volume is approximately $G=(V,E)$20 lower than the best Zoltan method; on 2.5D and 3D meshes, it still improves total communication volume by $G=(V,E)$21–$G=(V,E)$22; SpMV communication time is lowest for Geographer in all classes; edge cut is within $G=(V,E)$23–$G=(V,E)$24 of the best competitor; and, for a $G=(V,E)$25 billion-node Delaunay mesh with $G=(V,E)$26, Geographer takes approximately $G=(V,E)$27 whereas Zoltan–MJ takes $G=(V,E)$28 but gives $G=(V,E)$29–$G=(V,E)$30 worse communication volume. The method is also reported to exhibit near-perfect weak scaling to approximately $G=(V,E)$31 ranks and graceful degradation up to $G=(V,E)$32 (Looz et al., 2018).

6. Evaluation protocols, limitations, and extensions

The three approaches are evaluated with markedly different protocols. For non-rigid mesh clustering, the stated experimental blueprint proposes standard $G=(V,E)$33-mesh benchmarks such as FAUST and TOSCA, evaluation by clustering accuracy after best-permutation matching, normalized mutual information, and adjusted Rand index, comparison against standard K-means on raw vertices, iterative closest point plus K-means, and spectral embedding plus K-means, and qualitative inspection of learned centroids $G=(V,E)$34 and per-cluster warp fields $G=(V,E)$35. The same blueprint specifies the expected outcome that ST K-Means separates identity/pose classes more cleanly than rigid-only baselines (Cosentino et al., 2022).

For PageRank-based mesh segmentation, the evidence presented is qualitative: visually coherent regions, sharp boundaries at concavities, and convergence in a small number of iterations, but no quantitative “purity” or “IoU” numbers (Hajij et al., 2020). For balanced geometric partitioning, evaluation is explicitly system-oriented, with communication volume, edge cut, block diameter, SpMV communication time, and partitioning time as the primary metrics (Looz et al., 2018).

The limitations are also method-specific. Spatial Transformer K-Means is stated to have nonconvexity, scalability issues when $G=(V,E)$36, $G=(V,E)$37, and $G=(V,E)$38 are large, and a risk of over-warping when $G=(V,E)$39 or $G=(V,E)$40 are mis-set, so that warps may “cheat” and collapse distinct shapes. Its listed extensions are hierarchical ST K-Means, replacing TPS with deformation-graph or neural ODE diffeomorphisms, multi-scale coarse-to-fine landmark grids, and graph-smoothing regularizers in $G=(V,E)$41 (Cosentino et al., 2022). The PageRank formulation does not provide an explicit single objective function and does not report quantitative mesh-segmentation scores (Hajij et al., 2020). Balanced $G=(V,E)$42-means notes that the SFC-based global sort and subsequent redistribution can dominate when $G=(V,E)$43, and that extremely nonuniform point distributions require careful tuning of the influence update rate and erosion (Looz et al., 2018).

A recurrent misconception is that mesh clustering requires an arithmetic centroid in the ambient coordinate space. The graph formulation explicitly rejects that premise because, on a general graph or mesh, there is no vector mean; the non-rigid formulation shows that centroids can instead be defined after alignment in the warped space; and the balanced partitioning formulation treats the centroid as only one component of a constrained partitioning objective. This suggests that “K-means cluster meshes” is best understood not as a single algorithm, but as a class of Lloyd-type procedures whose mesh awareness is introduced through deformation invariance, geodesic structure, graph centrality, or balance constraints.