UDF-GMA: Dual Perspectives in Clinical and Geometric Domains

• UDF-GMA is an acronym with dual definitions: one for a Bayesian deep learning approach in clinical general movement assessment and another for medial-axis extraction in UDF mesh reconstruction.
• In the clinical setting, UDF-GMA fuses epistemic and aleatoric uncertainties via MC Dropout to predict poor repertoire in infant pose sequences with enhanced reliability.
• In geometry, UDF-GMA reformulates UDF 0-level set extraction as medial-axis extraction of an α-offset volume, enabling robust mesh reconstruction without sign-propagation.

UDF-GMA is an acronym used in two distinct ways in the supplied arXiv literature. In "UDF-GMA: Uncertainty Disentanglement and Fusion for General Movement Assessment" it denotes a pose-based automated GMA framework that explicitly models epistemic uncertainty in model parameters and aleatoric uncertainty from data noise, and then fuses these uncertainties with an embedded motion representation to predict poor repertoire from infant pose sequences (Luo et al., 7 Jul 2025). In the detailed presentation associated with "OffsetAxis: UDF Mesh Reconstruction via Offset-Volume Medial Axis Extraction," the same acronym is used for an "Unsigned Distance Field – Generalized Medial Axis" formulation that restates UDF $0$-level set extraction as medial-axis extraction of an $\alpha$-offset volume (Huang et al., 14 May 2026). The shared acronym therefore spans Bayesian deep learning for clinical movement analysis and geometry processing for mesh reconstruction.

1. Terminological scope and disambiguation

The duplication of the acronym is a substantive feature of the record rather than a typographic coincidence. One usage is clinical and probabilistic; the other is geometric and topological.

Expansion	Domain	Core task
Uncertainty Disentanglement and Fusion for General Movement Assessment	Pose-based automated GMA	Predicting poor repertoire
Unsigned Distance Field – Generalized Medial Axis	UDF mesh reconstruction	$0$-level set extraction via medial axis of the $\alpha$-offset volume

This suggests that unambiguous citation requires the full title or arXiv identifier. Without that disambiguation, the acronym alone is insufficiently specific for scholarly reference (Luo et al., 7 Jul 2025, Huang et al., 14 May 2026).

2. UDF-GMA for General Movement Assessment: clinical setting and uncertainty model

General Movement Assessment is an established clinical tool for evaluating the neurological integrity of infants by qualitatively inspecting their spontaneous whole-body movements—arms, legs, trunk, and neck—at preterm, writhing ($0$–$2$ months corrected age), and fidgety ($2$–$5$ months corrected age) stages. Early "Poor Repertoire" movements in preterm and writhing periods are known predictors of later neurodevelopmental impairment, including cerebral palsy. Automating GMA with video- or pose-based deep learning is motivated by the scarcity of expert assessors, but the approach is hampered by two fundamental forms of uncertainty: aleatoric uncertainty arising from noise in 2D/3D pose estimation, and epistemic uncertainty due to limited training data and lack of knowledge about model parameters (Luo et al., 7 Jul 2025).

The probabilistic formalization is cast in Bayesian neural network terms. Network weights $\omega$ are treated as random variables with posterior $p(\omega \mid \mathcal{D})$ for training data $\alpha$0. For a test pose sequence $\alpha$1, the predictive distribution is

$\alpha$2

The total predictive variance is decomposed into an epistemic term, $\alpha$3, and an aleatoric term, $\alpha$4. Because exact Bayesian inference is intractable for deep nets, the method employs MC Dropout as an approximate posterior $\alpha$5. With dropout-thinned weights $\alpha$6, logits $\alpha$7, and probabilities $\alpha$8, the predictive mean is $\alpha$9, and the epistemic uncertainty estimate is $0$0.

A central claim of the framework is that only by modeling both uncertainties can an automated system provide clinically reliable predictions, attenuating overconfident errors on noisy or out-of-distribution inputs and flagging high-uncertainty cases for expert review. A common misconception is that uncertainty in automated GMA is a single scalar nuisance; the formulation explicitly rejects that view by separating data uncertainty from model uncertainty.

3. Architectural formulation and objective in the GMA variant

The input is a 2D pose sequence $0$1 with dimensions $0$2. A compact motion embedding $0$3 is first extracted via CTR-GCN, identified as a state-of-the-art spatio-temporal graph convolutional backbone. UDF-GMA then introduces an Uncertainty Disentanglement Module and an Uncertainty-Guided Fusion Module (Luo et al., 7 Jul 2025).

Within the Uncertainty Disentanglement Module, the branch $0$4 is a small MLP with Dropout that outputs a scalar logit $0$5. Applying MC Dropout at inference yields $0$6 and, via the sigmoid-based estimates above, the epistemic uncertainty $0$7. The branch $0$8 is an MLP without Dropout, ending in a Softplus that produces the aleatoric estimate $0$9. The two uncertainty quantities are then combined through an MLP $\alpha$0 to obtain total predictive variance $\alpha$1.

Aleatoric uncertainty is trained through a heteroscedastic formulation that assumes the final logit follows a Gaussian with data-dependent variance:

$\alpha$2

where

$\alpha$3

Using $\alpha$4 samples $\alpha$5 with $\alpha$6, the average probability is

$\alpha$7

and the uncertainty loss is

$\alpha$8

The penalty term is designed so that $\alpha$9 becomes large only when the data are indeed noisy.

The Uncertainty-Guided Fusion Module injects the disentangled uncertainties into the motion representation. Epistemic fusion is defined by

$0$0

while aleatoric fusion is

$0$1

The final representation is

$0$2

which is passed to a classification head $0$3 to produce

$0$4

Training combines the standard video-level binary cross-entropy loss

$0$5

with the uncertainty regularizer:

$0$6

During inference, Dropout remains active only in $0$7 to estimate $0$8; all other Dropout layers are disabled.

4. Experimental profile of the GMA variant

The experimental benchmark is Pmi-GMA, consisting of 1120 thirty-second clips from 87 preterm infants, with 567 PR and 553 normal clips, captured at 10 FPS as 1280×720 overhead smartphone videos. Two evaluation protocols are reported. The intra-infant split uses 999 train, 111 val, and 121 test clips, with clips from the same infants appearing in all sets. The inter-infant split is stratified 65%/15%/20% by infant with no overlap. Five random seeds are used and mean±$0$9 is reported. Data augmentation, applied with 80% chance each sample, includes horizontal flip, time reversal, Gaussian noise, scaling, and magnitude/time warping. Evaluation metrics are Accuracy, Sensitivity, Specificity, and AUC-ROC. Baselines include FM-based methods—CA, STAM, WO-GMA, and smartphone-app—and WM-based methods—WMD with LDA, SVM, or RF, plus CTR-GCN representation (Luo et al., 7 Jul 2025).

On the inter-infant test, UDF-GMA reports ACC $2$0, SN $2$1, SP $2$2, and AUC $2$3. The next best method, WO-GMA, reports ACC $2$4, SN $2$5, SP $2$6, and AUC $2$7. The CTR-GCN baseline reports ACC $2$8, SN $2$9, SP $2$0, and AUC $2$1. On the intra-infant test, UDF-GMA reports ACC $2$2, SN $2$3, SP $2$4, and AUC $2$5—state-of-the-art.

Compared to the CTR-GCN baseline, the method yields $2$6 ACC, $2$7 SN, $2$8 SP, and $2$9 AUC on inter-infant evaluation. Low run-to-run variance is reported as evidence of robustness of uncertainty-guided learning. t-SNE visualizations of $5$0, $5$1 refined by [Ji ICCV ’23], and $5$2 show dramatically improved class separation only with UDF-fusion. Epistemic-threshold analysis further shows that by discarding test clips with $5$3 above a threshold $5$4, one can trade off data-retention versus sensitivity; at $5$5, SN reaches $5$6 while retaining approximately $5$7 of PR clips, enabling a high-sensitivity triage set for fully automated diagnosis while referring high-uncertainty cases to clinicians.

The limitations are explicit: reliance on 2D pose quality, no direct use of pose-estimation confidences, a single-center dataset of preterm/writhing GMs, and the computational cost of MC sampling. Future directions include validation on fidgety and longitudinal GMA datasets, incorporating keypoint confidence maps, investigating lighter uncertainty approximations such as Flipout and SWAG, and deploying a clinical decision-support tool that uses a $5$8 threshold to automatically accept high-confidence predictions while routing uncertain cases to expert review.

5. UDF-GMA as "Unsigned Distance Field – Generalized Medial Axis"

In the OffsetAxis presentation, UDF-GMA refers to a reconstruction formulation for unsigned distance fields rather than a clinical assessment model. The foundational object is a UDF

$5$9

defined by

$\omega$0

where $\omega$1 is the target shape. By definition, $\omega$2 everywhere and

$\omega$3

Unlike signed distance fields, UDFs make no inside/outside choice and can represent open or non-manifold geometries (Huang et al., 14 May 2026).

Given a scale $\omega$4, the $\omega$5-offset volume is

$\omega$6

with boundary

$\omega$7

If $\omega$8 is smaller than the shape’s weak feature size, then $\omega$9 is homotopy-equivalent to $p(\omega \mid \mathcal{D})$0. The medial axis of a volume,

$p(\omega \mid \mathcal{D})$1

becomes, for $p(\omega \mid \mathcal{D})$2, a thickened skeleton of the original shape. For small $p(\omega \mid \mathcal{D})$3, $p(\omega \mid \mathcal{D})$4 recovers $p(\omega \mid \mathcal{D})$5 up to homotopy and captures curves, junctions, and surface patches in one unified object.

The central reformulation is that the $p(\omega \mid \mathcal{D})$6-level set extraction problem can be restated as the extraction of the medial axis of the $p(\omega \mid \mathcal{D})$7-offset volume of the UDF. This avoids sign-propagation heuristics and addresses a common misconception that UDF extraction can simply reuse classical grid-based iso-surfacing. The OffsetAxis account states that classical grid-based iso-surfacing techniques are not applicable because they require a way to distinguish the inside from the outside of the shape.

6. Variational pipeline, reconstruction, and empirical behavior in the geometry variant

The geometry-oriented UDF-GMA approximates $p(\omega \mid \mathcal{D})$8 by a finite set of spheres $p(\omega \mid \mathcal{D})$9 and optimizes them against oriented offset-surface samples. For each sphere $\alpha$00 and its assigned cluster $\alpha$01 of samples $\alpha$02, the per-sphere energy is

$\alpha$03

where

$\alpha$04

is the Spherical Quadric Error Metric, and

$\alpha$05

is the line-quadric. Because both terms are quadratic in $\alpha$06, minimizing $\alpha$07 reduces to solving normal equations $\alpha$08 in closed form for $\alpha$09. To avoid degenerate or negative radii, the method enforces $\alpha$10 with $\alpha$11; if the free-radius solution violates this, $\alpha$12 is fixed to its previous value and only $\alpha$13 is updated by a $\alpha$14 linear system (Huang et al., 14 May 2026).

The pipeline has four main stages. First, offset-surface sampling is performed by ray casting and Poisson disk resampling: an expanded axis-aligned bounding box is built, random rays are cast, sphere tracing finds intersections with $\alpha$15, normals are estimated by $\alpha$16 or PCA on a $\alpha$17-NN graph filtered by gradient consistency, and Poisson disk resampling with radius $\alpha$18 enforces well-spaced samples. Second, medial sphere initialization forms candidate balls at $\alpha$19 with radius $\alpha$20, optionally refines them with a shrinking-ball algorithm, enlarges each radius to $\alpha$21, and greedily selects balls in descending $\alpha$22 order using graph-based flood fill. Third, a variational optimization loop alternates clustering and sphere updates until convergence or a maximum of 150 iterations. Fourth, mesh reconstruction builds a dual graph on sphere centers, inserts faces for 3-cliques, and applies topological thinning by removing simplex pairs in order of decreasing UDF-integral score. The remaining 2-simplices and 1-simplices form the final mesh, with triangles for surfaces and segments for curve branches.

The reported empirical profile spans DeepFashion, ShapeNetCar, 3DScene, and ABC + organics, with baselines including MC-variants such as CAP-UDF, MeshUDF, GeoUDF, and NSD-UDF, as well as DualMesh-UDF, DCUDF/DCUDF2, and MIND. Metrics include vertex count, Chamfer Distance, Hausdorff Distance, Triangle Quality, and, for Q-MDF models, Euler characteristic and non-manifold edge detection. The method is reported to achieve comparable or lower CD/HD than grid-based MC at $\alpha$23 with far fewer vertices, systematically higher triangle quality, correct recovery of non-manifold edges and Euler characteristic on Q-MDF, robustness to UDF noise of $\alpha$24 displacement, and support for direct point clouds and triangle soups. The stated limitations are that the choice of $\alpha$25 trades topology versus detail, and that initialization cost is $\alpha$26 for very dense samples. Future work includes adaptive sampling radii, accelerated covering, and formal topology guarantees.

The two meanings of UDF-GMA therefore occupy different technical ecosystems. One is a Bayesian uncertainty-aware classifier for pose-based automated GMA; the other is a medial-axis-based reconstruction strategy for unsigned distance fields. The shared acronym does not imply a shared methodology, but it does highlight the need for precise bibliographic disambiguation in arXiv-centered research practice.