Preconditioned Deformation Grids

• The paper introduces Preconditioned Deformation Grids as a method that directly estimates coherent deformation fields from unstructured point cloud sequences using multi-resolution voxel grids and Sobolev preconditioning.
• It achieves stable, drift-free dynamic reconstruction by coupling grid-based gradient optimization with a weak isometry term and confidence-weighted Chamfer loss.
• Neu-PiG extends the approach with a latent-grid encoding and time-modulated MLP, providing up to 60× faster convergence and consistent reconstructions on long sequences.

Searching arXiv for the cited papers to ground the article and confirm bibliographic details. Preconditioned Deformation Grids are a technique for dynamic surface reconstruction that estimates coherent deformation fields directly from unstructured point cloud sequences without requiring or forming explicit correspondences. The method represents motion with multi-resolution voxel grids and couples this representation to grid-based Sobolev preconditioning inside gradient-based optimization, so that a Chamfer loss between the input point clouds and an evolving template mesh, complemented by a weak isometry term on mesh edges, is sufficient to obtain accurate deformations (Kaltheuner et al., 22 Sep 2025). In subsequent work, the same core idea was reformulated as a neural preconditioned latent-grid encoding in Neu-PiG, which parameterizes the deformation of an entire long sequence relative to a single keyframe surface and decodes per-frame 6-DoF deformations with a lightweight MLP while retaining Sobolev-preconditioned optimization (Kaltheuner et al., 25 Feb 2026).

1. Problem setting and conceptual scope

Dynamic surface reconstruction of objects from point cloud sequences is presented as a challenging field in computer graphics. The central difficulty is to recover temporally coherent surfaces from unstructured observations while avoiding over-smoothing, poor generalization to unseen objects and motions, or optimization drift over long sequences. Preconditioned Deformation Grids were introduced to address these limitations without depending on multiple regularization terms or extensive training data (Kaltheuner et al., 22 Sep 2025).

The method is explicitly positioned against two classes of alternatives. One class relies on incremental deformation optimization, which risks drift and requires long runtimes on very long sequences. The other relies on complex learned models that demand category-specific training. Neu-PiG states these limitations directly and proposes a fast deformation optimization method that optimizes “from scratch,” while PDG emphasizes direct estimation from unstructured point cloud sequences without explicit correspondences (Kaltheuner et al., 25 Feb 2026).

A common misconception is that coherent deformation estimation necessarily requires explicit correspondences or category-specific pretraining. The cited works reject this premise: PDG estimates deformation fields directly from unstructured point cloud sequences without requiring or forming explicit correspondences, and Neu-PiG states that it completely avoids the need for any explicit correspondences or further priors (Kaltheuner et al., 22 Sep 2025).

2. Multi-resolution deformation parameterization

In Preconditioned Deformation Grids, the deformation field is represented by a hierarchy of voxel grids

$$G_t = \{ G_t^1, G_t^2, \dots, G_t^l, \dots, G_t^L \},$$

with each level $l$ covering the normalized domain $D=[-1,1]^3$ by a regular lattice of cells $C^l$. Level $l$ has a cell-spacing $\Delta^l \propto (2l-1)^{-1}$, and each cell $c\in C^l$ stores a 6D transformation parameter

$$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$

where $z\in\mathbb R^3$ parameterizes rotation via the Cayley map and $t\in\mathbb R^3$ is translation (Kaltheuner et al., 22 Sep 2025).

To deform a point $l$0 at time $l$1, the method gathers the eight trilinear weights from the enclosing cell at each level and averages over all levels: $l$2 Here $l$3 denotes the eight neighboring cells in level $l$4, and the factor $l$5 ensures equal contribution from each scale. This design captures overall motion at varying spatial scales and provides a flexible deformation representation (Kaltheuner et al., 22 Sep 2025).

Neu-PiG preserves the multi-resolution grid principle but replaces per-time-step transformation grids with a latent-grid encoding tied to a keyframe surface. It assumes a fixed reference mesh at keyframe $l$6 with vertices $l$7 and normals $l$8. The method stores two voxel-grid hierarchies: a position grid $l$9 with $D=[-1,1]^3$0 levels, where level $D=[-1,1]^3$1 has resolution $D=[-1,1]^3$2 and each cell stores a learnable 30-D feature, and a normal grid $D=[-1,1]^3$3 of fixed resolution $D=[-1,1]^3$4, where each cell stores a 2-D feature (Kaltheuner et al., 25 Feb 2026).

For a reference vertex, trilinear interpolation is performed in both the position and normal grids. The position features are averaged across levels to obtain $D=[-1,1]^3$5, while the normal feature gives $D=[-1,1]^3$6. Each reference vertex is thereby associated with a 32-D latent

$D=[-1,1]^3$7

This suggests a shift from explicit per-frame grid transforms to a surface-conditioned latent field that encodes entire deformations across all time steps (Kaltheuner et al., 25 Feb 2026).

3. Sobolev preconditioning and its role in optimization

The defining technical feature of Preconditioned Deformation Grids is the use of Sobolev preconditioning in the optimization loop. The continuous formulation introduces the $D=[-1,1]^3$8 inner product for scalar fields $D=[-1,1]^3$9 on $C^l$0: $C^l$1 For $C^l$2, this recovers a penalty on function value and gradient. The discrete version is built from a sparse graph Laplacian $C^l$3 defined on the voxel adjacency graph at each grid level (Kaltheuner et al., 22 Sep 2025).

At level $C^l$4, all transform components are stacked into a vector $C^l$5. A discrete Sobolev inner product is written as

$C^l$6

with

$C^l$7

as the preconditioner matrix. Given a loss $C^l$8 and gradient $C^l$9, the Sobolev-preconditioned descent direction is

$l$0

In practice, the paper uses the symmetric form $l$1 inside each update: $l$2 Because $l$3 is fixed, $l$4 or $l$5 can be applied by a sparse-linear solve at each iteration (Kaltheuner et al., 22 Sep 2025).

Neu-PiG transfers the same principle to latent-grid features. At each level $l$6, all cell features are stacked into $l$7, and with $l$8 the discrete Laplacian matrix on the voxel graph, one step of Sobolev-preconditioned gradient descent is

$l$9

Equivalently, with $\Delta^l \propto (2l-1)^{-1}$0,

$\Delta^l \propto (2l-1)^{-1}$1

The implementation uses two successive sparse solves with conjugate-gradient, or a small number of Jacobi/Gauss–Seidel iterations (Kaltheuner et al., 25 Feb 2026).

The stated purpose of preconditioning is not merely regularization in the conventional sense. Unpreconditioned grid optimization treats each cell’s latent update independently, which is reported to cause slow convergence, high-frequency artifacts, and drift over time. By contrast, the operator $\Delta^l \propto (2l-1)^{-1}$2 directly low-pass filters the gradient each step, couples each cell with its immediate neighbors, suppresses high-frequency noise at the source, accelerates convergence, and eliminates drift. Neu-PiG reports convergence that is often 5–10× faster per epoch and stable reconstructions on sequences of 100+ frames (Kaltheuner et al., 25 Feb 2026).

4. Objective functions and optimization pipeline

The PDG objective couples data fitting and weak geometric regularity. Given two point sets $\Delta^l \propto (2l-1)^{-1}$3, the squared Chamfer distance is

$\Delta^l \propto (2l-1)^{-1}$4

An initial template mesh $\Delta^l \propto (2l-1)^{-1}$5 is maintained and deformed through cumulative transforms to obtain $\Delta^l \propto (2l-1)^{-1}$6. The transform fitting term is

$\Delta^l \propto (2l-1)^{-1}$7

The method also defines a weak isometry loss over the edge set $\Delta^l \propto (2l-1)^{-1}$8 of the reference mesh: $\Delta^l \propto (2l-1)^{-1}$9 which softly enforces preservation of intrinsic edge lengths. The full objective is

$c\in C^l$0

with $c\in C^l$1, chosen so that $c\in C^l$2 contributes only a weak ($c\in C^l$3) regularization (Kaltheuner et al., 22 Sep 2025).

The optimization is simultaneous over template-mesh vertices $c\in C^l$4 and voxel-grid transforms $c\in C^l$5 for all $c\in C^l$6 and $c\in C^l$7. The mesh is preconditioned with its own Laplacian and $c\in C^l$8, with learning rate $c\in C^l$9. The grid uses a base learning rate $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$0, strengthened per finer level by a factor $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$1, while the smoothing weight is

$$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$2

All levels are updated in parallel via Adam plus preconditioning (Kaltheuner et al., 22 Sep 2025).

To prevent drift over long sequences, PDG includes a confidence-weighted Chamfer term: $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$3 where $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$4 over epochs so that later frames gradually regain full weight. The implementation description further specifies normalization of input points and mesh to $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$5, keyframe selection by maximizing occupied voxels near the temporal midpoint, reconstruction of $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$6 via screened Poisson, pruning of inactive cells, and output as a temporally coherent mesh sequence $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$7 (Kaltheuner et al., 22 Sep 2025).

Neu-PiG retains a two-term loss. Let $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$8 be the deformed mesh at time $$T_t^{l,c} = [ z_t^{l,c} ; t_t^{l,c} ] \in \mathbb R^6,$$9 and $z\in\mathbb R^3$0 the input point cloud. The deformation term is

$z\in\mathbb R^3$1

and the total loss is

$z\in\mathbb R^3$2

with $z\in\mathbb R^3$3 so that $z\in\mathbb R^3$4 contributes roughly 10% of $z\in\mathbb R^3$5. The isometry term is stated to preserve local shape and prevent folding (Kaltheuner et al., 25 Feb 2026).

5. Neu-PiG as a neural reformulation of preconditioned grids

Neu-PiG can be understood as a neural reformulation of the preconditioned-grid idea for long sequences. Rather than storing a separate 6D transform field for every frame and scale, it encodes entire deformations across all time steps at various spatial scales into a multi-resolution latent grid parameterized by the position and normal direction of a reference surface from a single keyframe. This latent representation is then augmented for time modulation and decoded into per-frame 6-DoF deformations via a lightweight MLP (Kaltheuner et al., 25 Feb 2026).

For frame $z\in\mathbb R^3$6, Neu-PiG computes a Fourier time embedding, with normalized time $z\in\mathbb R^3$7 and

$z\in\mathbb R^3$8

The decoder input for vertex $z\in\mathbb R^3$9 is

$t\in\mathbb R^3$0

This 40-D vector is fed into a shallow MLP $t\in\mathbb R^3$1 with three fully-connected layers of width 512 and LeakyReLU activations (Kaltheuner et al., 25 Feb 2026).

The final linear layer outputs a 7-D vector $t\in\mathbb R^3$2. The component $t\in\mathbb R^3$3 parameterizes rotation via a quaternion offset $t\in\mathbb R^3$4 and unit-normalization, while $t\in\mathbb R^3$5 is passed through $t\in\mathbb R^3$6 to bound translations. The resulting rigid transform $t\in\mathbb R^3$7 is applied to $t\in\mathbb R^3$8 to yield $t\in\mathbb R^3$9 (Kaltheuner et al., 25 Feb 2026).

This suggests that Neu-PiG preserves the optimization-centered character of PDG while compressing the spatio-temporal deformation field into a single latent representation anchored to a reference surface. The paper’s explicit comparison to PDG’s per-frame preconditioning further indicates that the principal innovation is not the abandonment of preconditioning, but its relocation from transform parameters to a unified latent grid over all time steps (Kaltheuner et al., 25 Feb 2026).

6. Empirical profile, reported gains, and interpretation

The empirical claims reported for PDG and Neu-PiG emphasize both fidelity and runtime, especially on long sequences. PDG states that extensive evaluations demonstrate superior results, particularly for long sequences, compared to state-of-the-art techniques. Its implementation details report $l$00 by default, a coarse level $l$01 and level $l$02 roughly $l$03 with pruning of inactive cells, approximately 50% reduction in memory from pruning, GPU memory $l$04 on an RTX 4090 for a 17-frame sequence with $l$05 points/frame, and runtime $l$06 minutes for $l$07 (Kaltheuner et al., 22 Sep 2025).

Neu-PiG reports results on three standard benchmarks—DFAUST, AMA, and DT4D—and states that it outperforms all training-free baselines in both accuracy and runtime. The reported benchmark values are as follows (Kaltheuner et al., 25 Feb 2026):

Dataset	Baseline PDG Time	Neu-PiG Time
DFAUST	7 min	32 s
DT4D	7 min	32 s
AMA	7 min	32 s

For the same datasets, Neu-PiG reports Chamfer, NC, [email protected]%, and Corr. values: DFAUST with $l$08, $l$09, $l$10, $l$11; DT4D with $l$12, $l$13, $l$14, $l$15; and AMA with $l$16, $l$17, $l$18, $l$19 (Kaltheuner et al., 25 Feb 2026).

The paper summarizes these results by stating that Neu-PiG runs $l$20 faster than prior training-free optimizers, from 7 min to 32 s, and matches or exceeds the accuracy of category-specific learned methods such as M2V and CaDeX without any pretraining. It also states that inference speeds are on the same order as heavy pretrained models and that high-fidelity, drift-free surface reconstructions are obtained in seconds (Kaltheuner et al., 25 Feb 2026).

A useful synopsis of the method family is:

Aspect	PDG	Neu-PiG
Core representation	Multi-resolution voxel grids with 6D cell transforms	Position and normal latent grids on a keyframe surface
Optimization variable	$l$21 and template mesh $l$22	Latent voxel features and decoder weights
Temporal strategy	Cumulative transforms and confidence-weighted Chamfer	Unified latent grid with time-modulated MLP

The principal interpretation supported by the cited material is that preconditioning is the organizing idea across both methods. In PDG, it structures optimization over explicit deformation grids; in Neu-PiG, it structures optimization over latent grids that encode entire long sequences. A plausible implication is that the reported gains in speed and stability are tied less to any single loss term than to the combination of multi-scale spatial encoding with Sobolev-filtered gradient updates (Kaltheuner et al., 22 Sep 2025).

7. Relation between PDG and Neu-PiG

The relationship between the two papers is cumulative rather than discontinuous. Preconditioned Deformation Grids introduced the core ingredients: multi-resolution voxel grids, Sobolev preconditioning applied per grid level, Chamfer-based fitting to point clouds, a weak isometry prior on mesh edges, and a confidence-weighted mechanism to prevent drift over long sequences (Kaltheuner et al., 22 Sep 2025).

Neu-PiG explicitly states that it gives a focused, end-to-end technical description of its core, the neural preconditioned deformation grids, beginning with the multi-resolution latent grid, deriving the Sobolev preconditioner used during training, writing out the reconstruction and isometry losses, describing the time-modulated MLP decoder in detail, and explaining why preconditioning yields fast, drift-free convergence. It then closes with a concise summary of empirical speedups and fidelity gains (Kaltheuner et al., 25 Feb 2026).

The later method therefore preserves the same foundational commitments—optimization from scratch, no explicit correspondences, spatial smoothness induced through Sobolev operators, and temporally consistent reconstruction from point cloud sequences—while altering the representation and decoder. Compared to PDG’s per-frame preconditioning, Neu-PiG’s single, unified latent grid enforces smoothness across all time steps, which the paper associates with stable reconstructions on sequences of 100+ frames (Kaltheuner et al., 25 Feb 2026).

Within this lineage, “preconditioned deformation grids” denotes both a specific 2025 method and a broader methodological template for dynamic surface reconstruction: represent deformations on multi-scale grids, optimize them directly from geometric losses, and shape the optimization trajectory by applying Sobolev structure to the gradient rather than relying on explicit correspondences or category-specific pretraining (Kaltheuner et al., 22 Sep 2025).