GUFU: Reconstruction & Localization Methods

• GUFU is an acronym for two distinct methods: one using 2D Gaussian planes with implicit UDFs for open-surface reconstruction, and another using graph-based fingerprint updates in indoor localization.
• The GaussianUDF approach replaces volumetric Gaussians with thin 2D planes, combining differentiable alpha blending and Chamfer-distance supervision to accurately capture challenging open surfaces.
• The WiFi GUFU method employs an autoencoder and GraphSAGE-like GNN to update stale radio maps through virtual edges and link prediction, significantly reducing RSS and location errors.

Searching arXiv for papers using “GUFU” and closely related terms to ground the article. GUFU is an acronym with at least two distinct uses in recent arXiv literature. In geometric reconstruction, it denotes GaussianUDF, a method for reconstructing open surfaces from multi-view images by coupling 2D Gaussian planes with unsigned distance functions (UDFs) and 3D Gaussian Splatting (3DGS) (Li et al., 25 Mar 2025). In indoor localization, it denotes Graph-based Fingerprint Update Using Unlabelled WiFi Signals, a graph-based system for updating WiFi fingerprint databases from unlabeled crowdsourced measurements under access-point (AP) churn (Chiu et al., 15 Jul 2025). The label should also be distinguished from GUF, which refers elsewhere to the Galam Unifying Frame and to the Geometric Unscented Kalman Filter, and from gUFO, an OWL 2 DL implementation of the Unified Foundational Ontology (Galam et al., 2013, Fang et al., 2020, Almeida et al., 21 Mar 2026).

1. Acronym scope and disambiguation

The acronym is not associated with a single research lineage. It is reused across computer vision and wireless localization, while adjacent strings such as GUF and gUFO denote unrelated constructs in statistical physics, nonlinear filtering, and Semantic Web ontology engineering.

Label	Expansion	Domain
GUFU	GaussianUDF	open-surface reconstruction from multi-view images
GUFU	Graph-based Fingerprint Update Using Unlabelled WiFi Signals	indoor WiFi fingerprint database maintenance
GUF	Galam Unifying Frame	two-state dynamics and the 2D Ising transition
GUF	Geometric Unscented Kalman Filter	nonlinear Bayesian state estimation
gUFO	A Gentle Foundational Ontology for Semantic Web Knowledge Graphs	OWL 2 DL foundational ontology

The two GUFU usages are method names rather than variants of a common framework. One addresses the mismatch between explicit Gaussian splats and continuous implicit distance fields; the other addresses the mismatch between a stale radio map and unlabeled RSS batches containing both shared and new APs.

2. GaussianUDF: UDF learning through 2D Gaussian planes

GaussianUDF defines GUFU as a method for reconstructing open surfaces from multi-view images by combining the speed and explicit geometry of 3DGS with the flexibility of UDFs. The motivation is specific to open surfaces such as clothing, leaves, and curtains, for which an SDF-style inside/outside partition is problematic. The UDF is written as

$d = f(\mathbf{q}), \qquad \mathbf{q}=(x,y,z),$

with zero level set

$\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$

The gradient $\nabla f(\mathbf{q})$ points away from the nearest surface and is used to pull points toward the surface (Li et al., 25 Mar 2025).

The method’s central representational move is to replace ordinary 3D Gaussians with thin, flat 2D Gaussian planes. Each Gaussian $\mathbf{g}_i$ is parameterized by a center $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$, color $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$, opacity $\alpha_i$, rotation matrix $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$, and scaling factors $\mathbf{s}_i \in \mathbb{R}^{1\times 2}$. The last column of $\mathbf{r}_i$ gives the plane normal $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$0. Because the support is 2D rather than volumetric, the Gaussian behaves as a local surfel or surface patch, which the paper argues captures open surfaces and boundaries better than volumetric or isotropic Gaussian blobs.

Rendering is performed by differentiable alpha blending,

$\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$1

with the same pipeline also producing depth maps and normal maps. The appearance supervision uses

$\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$2

together with the standard SSIM-style image quality term $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$3. In this construction, the Gaussian planes act as explicit surface carriers, while the UDF supplies a continuous implicit field over $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$4.

3. GaussianUDF optimization, supervision, and empirical behavior

GaussianUDF jointly optimizes the UDF $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$5 and the 2D Gaussian planes $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$6 through a coupled objective. Far-region supervision is generated by sampling query points $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$7 around Gaussian centers using the Neural-Pull sampling strategy and projecting them toward the zero level set via

$\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$8

The projected points are matched to Gaussian centers by a Chamfer-distance-based $\mathcal{S} = \{\mathbf{q}\mid f(\mathbf{q})=0\}.$9. Near-region supervision is denser: for root points $\nabla f(\mathbf{q})$0 sampled on a Gaussian plane and offsets $\nabla f(\mathbf{q})$1, points

$\nabla f(\mathbf{q})$2

are assigned ground-truth unsigned distance $\nabla f(\mathbf{q})$3, yielding

$\nabla f(\mathbf{q})$4

The paper treats this near-plane self-supervision as essential because Gaussian centers alone are sparse and noisy.

Optimization around the zero level set is explicitly stabilized. Gaussian centers are projected to the zero level set, gradients through $\nabla f(\mathbf{q})$5 are stopped, and the centers are updated with

$\nabla f(\mathbf{q})$6

Two additional regularizers are used: the 2DGS depth distortion loss

$\nabla f(\mathbf{q})$7

and a normal consistency term

$\nabla f(\mathbf{q})$8

where $\nabla f(\mathbf{q})$9 is estimated from rendered depth gradients. The full objective is

$\mathbf{g}_i$0

The reported evaluation covers DF3D, DTU, two public NeUDF real-captured scenes, and four author-captured real scenes. Training uses 30k iterations, an 8-layer MLP with 256 hidden units and ReLU, absolute-value last-layer activation, positional encoding, initial learning rate $\mathbf{g}_i$1 with cosine decay, a single NVIDIA 3090, 500 Gaussian planes per batch, 10 root points per plane, $\mathbf{g}_i$2 on DF3D, and $\mathbf{g}_i$3 on DTU. On DF3D, the reported average Chamfer Distance is 1.60 for GUFU, compared with 1.71 for VRPrior, 1.98 for 2S-UDF, 2.15 for NeuralUDF, 2.49 for GOF, 3.81 for 2DGS, and 4.36 for NeuS; training time is about 1.6h. On DTU, the reported average CD is 0.68, compared with 0.71 for VRPrior, 0.74 for GOF, 0.75 for GSPull, and 0.80 for 2DGS. The ablation on DTU reports CD $\mathbf{g}_i$4 for only $\mathbf{g}_i$5, $\mathbf{g}_i$6 for Far + Near, $\mathbf{g}_i$7 for Far + Proj, $\mathbf{g}_i$8 when removing warp loss, $\mathbf{g}_i$9 when removing Near, $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$0 when removing Proj, and $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$1 for the full model (Li et al., 25 Mar 2025).

4. WiFi GUFU: graph formulation for unlabeled fingerprint updates

In indoor localization, GUFU expands to Graph-based Fingerprint Update using Unlabeled WiFi signals. The method addresses the case in which an initial fingerprint database,

$\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$2

must be updated from an unlabeled batch

$\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$3

even though the AP set may have changed. The target is not append-only augmentation but amendment of the RSS values at existing designated locations. Missing AP observations are filled with $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$4 dBm, then normalized to $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$5 by adding 120 and dividing by 120. An autoencoder with encoder $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$6 and decoder $\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$7 produces compact embeddings

$\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$8

trained by reconstruction loss

$\mathbf{\mu}_i \in \mathbb{R}^{1\times 3}$9

These embeddings become node features in a heterogeneous weighted graph (Chiu et al., 15 Jul 2025).

The graph is

$\mathbf{c}_i \in \mathbb{R}^{1\times 3}$0

with two node types: sample nodes $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$1 and AP nodes $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$2. If AP $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$3 is detected in sample $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$4, GUFU creates an edge $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$5 with shifted RSS weight

$\mathbf{c}_i \in \mathbb{R}^{1\times 3}$6

where $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$7. Sample-node features are encoder outputs augmented with the location label $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$8 during offline training. AP-node features are initialized as weighted averages of neighboring sample features. A distinctive addition is the use of virtual edges between sample nodes whose cosine similarity exceeds a threshold $\mathbf{c}_i \in \mathbb{R}^{1\times 3}$9; the reported setting is $\alpha_i$0. The stated rationale is that thresholding avoids forcing in weakly similar neighbors, unlike top-$\alpha_i$1 selection.

The GNN is GraphSAGE-like and uses both node and edge features. For each edge $\alpha_i$2, the initial edge feature is

$\alpha_i$3

Node embeddings are updated by aggregating concatenated neighbor-node and edge features, followed by $\alpha_i$4 normalization, and edge features are updated analogously. The graph loss is

$\alpha_i$5

After offline training, virtual edges are discarded to save space.

5. WiFi GUFU: update module, AP churn, and reported performance

When an unlabeled batch arrives, GUFU first inserts new sample nodes, encodes them as

$\alpha_i$6

and recreates virtual edges among old and new sample nodes using the same cosine-similarity threshold. It then applies an update module with two sequential MLPs: a location predictor that maps new sample embeddings to predicted location labels, and a fingerprint updater that maps location labels from the existing database to updated feature vectors for the existing fingerprints. The decoded output $\alpha_i$7 replaces the stale RSS values. The MLP losses are

$\alpha_i$8

with neighborhood consistency terms

$\alpha_i$9

and final mixture

$\mathbf{r}_i \in \mathbb{R}^{3\times 3}$0

where $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$1.

AP churn is handled explicitly. For newly observed APs, GUFU creates new AP nodes, initializes their features by weighted averaging over neighboring sample nodes, and predicts missing edges using goodness and fairness vectors: $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$2 For a candidate pair $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$3, the predicted edge weight is

$\mathbf{r}_i \in \mathbb{R}^{3\times 3}$4

If $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$5, the edge is added; if $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$6 and the edge exists, it is removed. The threshold is

$\mathbf{r}_i \in \mathbb{R}^{3\times 3}$7

If an AP node loses all its edges, it is deleted from the graph. Incremental retraining updates graph structure, reruns GNN inference, retrains the GNN, and retrains the autoencoder with an augmented loss that preserves both reconstruction and alignment.

The reported evaluation uses four real sites: a campus building, Mall A, Mall B, and a hospital. GUFU is compared with Fidora, iToLoc, MTDAN, and WiDAGCN using RSS error and location error. The paper reports an overall 21.4% reduction in RSS error and a 29.8% reduction in location prediction error. Reported average location errors are 4.92 m on Campus, 4.44 m on Mall A, 4.39 m on Mall B, and 3.06 m on Hospital. The ablations report that removing the autoencoder feature extractor increases RSS error by up to 26.6% and location error by up to 18.9%; removing virtual edges worsens RSS error by 11.8% and location error by 15.2%; thresholding outperforms top-$\mathbf{r}_i \in \mathbb{R}^{3\times 3}$8 by 16.2% in RSS error and 22.1% in location error; and under up to 90% MAC removal, GUFU still achieves an average location error of 15.36 m, which is 46.2% better than GUFU without link prediction. The campus per-update times are reported as around 91–99 seconds. The stated limitations are periodic batch updates, sensitivity to the similarity threshold $\mathbf{r}_i \in \mathbb{R}^{3\times 3}$9, eventual need for a new site survey under large drift, designated-location updates only, and RSS-only input (Chiu et al., 15 Jul 2025).

6. Relation to adjacent terms and broader significance

The two GUFU methods are unrelated in formalism, data modality, and application domain. GaussianUDF is a differentiable reconstruction pipeline that uses 2D Gaussian planes as surface proxies and trains a continuous UDF through appearance fitting, far-field projection, near-plane self-supervision, and geometric regularization. The WiFi method is a graph-based maintenance system that uses an autoencoder, a heterogeneous graph with AP and sample nodes, virtual edges between similar samples, a GraphSAGE-style GNN with edge features, a two-MLP update module, and a goodness/fairness link-prediction mechanism to amend an evolving radio map. This suggests that GUFU functions as a field-local acronym rather than a transdisciplinary framework.

The distinction matters because GUF, without the terminal “U,” already has established meanings. In statistical physics, the Galam Unifying Frame is a coarse-grained recursion for the fraction $\mathbf{s}_i \in \mathbb{R}^{1\times 2}$0 of $\mathbf{s}_i \in \mathbb{R}^{1\times 2}$1 spins in a 5-spin local cluster and was used to study the 2D ferromagnetic Ising model, including an unexpected intermediate “dis/order” phase under the GUF–Metropolis combination (Galam et al., 2013). In nonlinear estimation, the Geometric Unscented Kalman Filter is a CKF-style Gaussian filter whose central innovation is geometric unscented sampling on ellipsoidal probability shells with positive weights (Fang et al., 2020). The visually similar gUFO is, again, different: it is a lightweight OWL 2 DL implementation of the Unified Foundational Ontology for Semantic Web knowledge graphs, with two disjoint taxonomies for individuals and types, reification patterns for qualities and relations, support for situations and high-order types, and SHACL/SPARQL constraints for ontological validation (Almeida et al., 21 Mar 2026).

Within current arXiv usage, GUFU therefore names two technically substantial but independent systems. One belongs to the reconstruction literature at the 3DGS–UDF interface; the other belongs to radio-map maintenance for indoor localization. Any interpretation of “GUFU” without disciplinary context is underspecified.