Smooth ARAP: Enhanced Mesh Deformation

• Smooth ARAP is a deformation framework that enhances classical ARAP with global smoothness constraints and robust numerical properties.
• It employs dense pairwise energy formulations and bi-Laplacian regularization to eliminate spiking artifacts while preserving local detail.
• The method achieves fast convergence and real-time performance, making it suitable for applications like medical imaging and interactive CAD.

Smooth As-Rigid-As-Possible (ARAP) frameworks represent recent advances in mesh and surface deformation methodologies designed to overcome major limitations of the classical ARAP model, particularly in terms of surface smoothness, global propagation of rigidity constraints, and continuity at manipulation handles. These approaches introduce either higher-order smoothness functionals or densely connected energy formulations that eliminate artifacts such as “spiking” at handles and enable globally coherent, real-time deformation with robust numerical properties and minimal loss of local detail (Zhou et al., 2020, Oehri et al., 17 Jan 2025).

1. Foundations of ARAP and the Need for Smoothness

The original ARAP energy provides a discrete deformation model ensuring that, locally, deformations are as close to rigid as possible while accommodating large, nonlinear global changes. ARAP is typically formulated on a mesh $M=(V,T)$ with cotangent weights $w_e$ on half-edges and area weights $A_v$ per vertex. The standard ARAP energy,

$$E_{\rm ARAP}(V',R)=\sum_{v\in V}A_v \sum_{e\in N(v)}\frac{w_e}{3A_v}\, \|\mathbf e' - R_v\,\mathbf e\|^2,$$

with local rotations $R_v$ minimizing per-vertex distortion, excels at preserving detail and can be minimized efficiently using local–global optimization. However, its restriction to immediate neighborhood (1-ring) edges induces two key problems: (1) slow propagation of smoothness constraints across the surface due to sparse coupling, and (2) the formation of spikes or C⁰ discontinuities, especially at point handles (Oehri et al., 17 Jan 2025).

2. Densely Connected ARAP for Global Smoothness

To address slow global propagation and surface artifacts, the dense ARAP (“smooth ARAP” in the context of deformable surface recovery) introduces fully connected, distance-weighted regularization among all sparse control-points on the surface. The deformation energy is

$$f_{\text{total}}(\{W_i^t\}) = f_{ICP} + w_{ORB}f_{ORB} + w_{ARAP}f_{ARAP},$$

where the smooth ARAP term is

$$f_{ARAP} = \sum_{i_1=1}^N \sum_{i_2=1}^N w_{i_1,i_2}\left( f_{\text{length},\,i_1i_2} + w_{\text{angle}}f_{\text{angle},\,i_1i_2} + w_{\text{rotation}}f_{\text{rotation},\,i_1i_2} \right).$$

Here, $f_{\text{length}}$ enforces length preservation, $f_{\text{angle}}$ penalizes angular deviations, and $f_{\text{rotation}}$ couples dual-quaternion rotations. The all-to-all structure, where $w_e$0 decays with initial Euclidean distance, ensures that even remote surface regions remain rigidly coordinated, enabling fast, low-iteration convergence to physically plausible surfaces (Zhou et al., 2020).

3. Higher Order (Bi-Laplacian) Regularization for $w_e$1 Continuity

A complementary approach to “smooth ARAP” modifies the key differential operator underpinning ARAP’s global step. By upgrading the original second-order (Poisson, Laplacian) operator to a fourth-order (bi-Laplacian) regularizer, smooth ARAP penalizes not just positional deviation but also discontinuities in surface derivatives. Specifically, in this framework,

$w_e$2

with $w_e$3 the area-weighted cotangent Laplacian vectors. The total energy blends ARAP and the smoothness penalty: $w_e$4 In the global solve, this system,

$w_e$5

enforces $w_e$6 continuity at handles, eliminating spikes and ensuring both positional and first derivative continuity (Oehri et al., 17 Jan 2025).

4. Implementation Paradigms and Optimization Strategies

Smooth ARAP schemes retain core implementation structures of the standard ARAP model while introducing new computational components. For densely connected ARAP, all pairwise residuals (length, angle, rotation) among $w_e$7 control-points are accumulated in parallel, exploiting GPU acceleration. At each frame, the Jacobian for each dual-quaternion warp $w_e$8 is constructed from the relevant ICP, ORB, and ARAP residuals; Levenberg–Marquardt updates are solved in parallel, with 5–10 iterations per frame being typically sufficient for convergence (Zhou et al., 2020).

For higher-order smooth ARAP, sparse Cholesky factorizations of the bi-Laplacian system matrix suffice for meshes with hundreds of thousands of vertices. The global linear system remains sparse (roughly twice as dense as the basic Laplacian); handle additions/removals are handled efficiently via KKT solves, requiring only small dense linear algebra per update and maintaining interactiveness ($w_e$90.11 s per global step for a 170k-vertex mesh) (Oehri et al., 17 Jan 2025).

5. Parameterization and Effects on Deformation Quality

Smooth ARAP frameworks introduce additional parameters that balance rigidity and smoothness:

Parameter	Typical Value / Range	Role
$A_v$0	20.0	Weights angular deviation penalty (dense ARAP)
$A_v$1	100.0	Strength of rotation coupling (dense ARAP)
$A_v$2	10.0	ORB feature weight (stereo video application)
$A_v$3	[0, 1) (e.g., 0.95)	ARAP $A_v$4 higher-order smooth blend

For dense ARAP, the global ARAP weight $A_v$5 is scaled dynamically per frame to maintain balanced influence relative to ICP/ORB data terms. The smooth ARAP parameter $A_v$6 smoothly interpolates between maximal detail preservation ($A_v$7) and strong smoothness ($A_v$8), with high values ($A_v$9) providing consistent smoothness with minimal parameter tuning (Zhou et al., 2020, Oehri et al., 17 Jan 2025).

6. Quantitative Performance and Practical Implications

Dense ARAP and higher-order smooth ARAP models achieve significant improvements in surface quality, numerical robustness, and real-time performance. In deformable surface recovery from stereo video, dense ARAP yields reconstruction RMSE $$E_{\rm ARAP}(V',R)=\sum_{v\in V}A_v \sum_{e\in N(v)}\frac{w_e}{3A_v}\, \|\mathbf e' - R_v\,\mathbf e\|^2,$$0 mm with only 5–8 LM iterations per frame, and maintains update rates of 15 Hz with sub-millimeter accuracy on challenging in vivo data (Zhou et al., 2020). In interactive shape modeling, higher-order smooth ARAP eliminates spiking artifacts at handles while reducing iteration counts compared to standard ARAP (e.g., 25 iterations vs. 294 for a 4,800 vertex mesh), and maintains real-time interactivity for meshes up to 170k vertices (Oehri et al., 17 Jan 2025).

A plausible implication is that smooth ARAP frameworks can be broadly applied in scenarios where both local detail and global smoothness are essential, such as surgical navigation, animation, and interactive CAD. These approaches preserve the favorable computational structure of ARAP, while extending it to high fidelity, artifact-free surface deformation.

7. Comparative Summary of Smooth ARAP Approaches

Method	Smoothness Strategy	Surface Quality	Computational Cost	Application Scenarios
Dense ARAP	Dense pairwise regularization	Sub-millimeter, globally smooth	Higher parallel workload	Real-time deformation (medical, vision)
Higher-Order ARAP	Bi-Laplacian regularization	$$E_{\rm ARAP}(V',R)=\sum_{v\in V}A_v \sum_{e\in N(v)}\frac{w_e}{3A_v}\, \|\mathbf e' - R_v\,\mathbf e\|^2,$$1 (tangent) continuity	Slightly denser matrices	Interactive shape modeling, animation

Both approaches overcome classical ARAP’s locality problem and handle artifacts without sacrificing detail or efficiency, and both are available in practical, real-time software pipelines (Zhou et al., 2020, Oehri et al., 17 Jan 2025).