Geometric Topology Fusion (GT Fusion)

• Geometric Topology Fusion is a family of methods that leverage inherent geometric and topological structures to fuse diverse data types in areas such as multi-sensor time series, cross-view geo-localization, 3D field theories, and volumetric reconstruction.
• Techniques include joint delay embeddings with Gram–Schmidt orthogonalization, cross-attention guided BEV integration, topological handle-slides in 3D theories, and non-manifold grid updates to accommodate topology changes.
• Empirical findings and theoretical analyses demonstrate enhanced manifold reconstruction, improved cross-view feature alignment, correct topological fusion in field theories, and robust handling of dynamic scene changes.

Geometric Topology Fusion (GT Fusion) appears in several distinct research settings rather than as a single universally fixed construction. In multi-sensor time-series analysis, it denotes a framework that combines a joint delay embedding with a controlled Gram–Schmidt orthogonalization in order to reconstruct a high-dimensional state space and correct geometric distortions induced by correlated observation functions (Solomon et al., 2020). In cross-view geo-localization, it denotes a module inside the CLGT framework that injects explicit Bird’s-Eye View (BEV) road-topology cues into street-view feature maps by local projection, overlapping spatial reduction, cross-attention, and Dual Dynamic Fusion (Ouyang et al., 13 Mar 2026). In the study of 3d theories $T[M_3]$, it denotes the connected-sum handle-slide of two Ooguri-Vafa matter-circles in a three-manifold $M_3$, with a field-theoretic realization through a vortex-partition-function fusion identity (Cheng, 2024). A related use of the phrase “geometric topology fusion” also appears in topology-change-aware volumetric reconstruction, where non-manifold volumetric grids permit connectivity updates by cell splitting and replication during 4D reconstruction (Li et al., 2020).

1. Terminological scope

The current literature associates GT Fusion with different mathematical objects, different fusion targets, and different technical aims. The commonality is not a shared algorithmic template, but the use of geometric or topological structure as the organizing principle for fusion.

Setting	Fused objects	Central mechanism
Multi-sensor time series	Delay embeddings from $m$ sensors	Joint delay embedding + Gram–Schmidt tensor
Cross-view geo-localization	Street-view and BEV feature maps	Cross-attention + Dual Dynamic Fusion
3d theories $T[M_3]$	Two matter circles in $M_3$	Connected-sum handle-slide + fusion identity
Dynamic scene reconstruction	TSDF and EDG connectivity under cuts	Non-manifold grid with cell splitting and replication

This distribution of meanings shows that the expression is domain-specific. In one case, topology enters through Takens-style reconstruction and persistence summaries; in another, through BEV road topology; in another, through surgery moves and plumbing graphs; and in another, through topological change in a canonical volume (Solomon et al., 2020).

2. Joint delay embeddings and orthogonalization in multi-sensor time series

In "Geometric Fusion via Joint Delay Embeddings," GT-Fusion is a two-step framework for fusing multi-sensor time series. The hidden state-space model is an unknown compact manifold $M$ of dimension $n$, equipped with a smooth flow or diffeomorphism $f:M\to M$. The sensors are observation functions $x_i:\mathbb R\to\mathcal O_i$, $i=1,\dots,m$, with a fixed delay $M_3$0 and embedding length $M_3$1. For each sensor,

$M_3$2

and the joint delay embedding is the combined time series $M_3$3 with matrix-entry notation

$M_3$4

Its reconstruction role is stated in terms of classical Takens and its extensions. If each $M_3$5 for generic observations $M_3$6, then for sufficiently large $M_3$7 the induced map into $M_3$8 is generically an embedding. Accordingly, $M_3$9 tracks a single orbit in $m$0, reconstructed in a high-dimensional space. The paper states that, from Takens’ theorem and its multivariate extension, the map $m$1 is a topological embedding of the underlying orbit in $m$2 when $m$3 is large enough, so the fusion preserves manifold structure and introduces no self-intersections or foldings by merely stacking delays (Solomon et al., 2020).

The second step is an orthogonalization scheme, the Gram–Schmidt tensor $m$4, introduced to mitigate tangencies between observation functions. Given two times $m$5, one forms

$m$6

collects $m$7, fixes an orthogonality parameter $m$8, repeatedly selects the currently largest unmarked vector, and partially projects the remaining vectors off that direction via

$m$9

The output is

$T[M_3]$0

When $T[M_3]$1, this is exactly Gram–Schmidt orthonormalization up to length ordering; when $T[M_3]$2, there is no correction. The fused distance is then

$T[M_3]$3

The geometric rationale is explicit. If two sensors are highly correlated, their delay-vectors are nearly parallel, and a naive Euclidean fusion over-counts redundant information and inflates the geometry along that common direction. The orthogonalization spreads such directions apart, enhances diversity, and lets each independent mode of variation contribute in proportion to its norm and angle relative to others. The paper states that, compared to naïve $T[M_3]$4 concatenation (JDL) or graph-based metric mixing (SNF), GT-Fusion yields a metric that more faithfully approximates the true Euclidean distance on $T[M_3]$5 (Solomon et al., 2020).

The theoretical and computational profile is also specified. Topological correctness follows from genericity assumptions in Takens’ theorem, and no additional convergence proof is provided for the geometry. Delay embedding construction costs $T[M_3]$6, per-pair orthogonalization costs $T[M_3]$7, and constructing a full $T[M_3]$8 fused distance matrix costs $T[M_3]$9. The paper notes that one often subsamples pairs or uses fast nearest-neighbor search to reduce cost.

Empirical evaluation is reported on both synthetic and real data. On a synthetic torus-curve benchmark with 100 points on a $M_3$0-curve on the torus in $M_3$1, the methods compared were JDL, SNF, and GT-Fusion with $M_3$2 and $M_3$3. The stated findings are that JDL and SNF severely distort the torus geometry, JDE with $M_3$4 already improves recovery, and JDE with full orthogonalization $M_3$5 and larger $M_3$6 best reconstructs the correct distance matrix. On the MotionSense dataset, using twelve smartphone-sensor channels and four tasks for a single subject, the reported qualitative summaries are that JDL produces noisy MDS embeddings and trivial persistence, JDE with $M_3$7 yields loops in “downstairs/walking” but often underestimates spherical or toric structure, and JDE with $M_3$8 consistently reveals clear 1-dimensional cycles and sometimes 2-dimensional features, including a sphere for the “downstairs” data (Solomon et al., 2020).

3. GT Fusion in cross-view geo-localization

In "CVGL: Causal Learning and Geometric Topology," GT Fusion is one of the two core modules in the CLGT framework. Its stated purpose is to inject explicit Bird’s-Eye View road-topology cues into street-view feature maps in order to close the semantic and geometric gap between ground-level panoramas and overhead imagery. The motivation is the extreme perspective difference between aerial images and street panoramas, together with occlusions and view-specific style variations; BEV representations are used because they encode the same underlying road-network topology in a canonical, view-invariant form (Ouyang et al., 13 Mar 2026).

The module takes two feature maps of identical spatial resolution $M_3$9 and channel dimension $M$0: $M$1 Here $M$2 is street-view features extracted by a ConvNeXt-B backbone after causal intervention, and $M$3 is BEV features obtained by first lifting the street panorama to BEV via geometric back-projection, as in HC-Net, and then encoding with the same ConvNeXt-B backbone with shared weights. The goal is to form a fused feature map $M$4 that enriches the street representation with BEV topology while preserving local street details.

Architecturally, GT Fusion is inserted after the final stage of the shared backbone, immediately before the Data-Adaptive Pooling module. Its data flow has four named stages. First, a depthwise $M$5 convolution called “proj” is applied to each branch: $M$6 Second, an overlapping spatial reduction downsamples to $M$7 with padding and overlap, and the result is added back and normalized: $M$8 Third, the module computes standard scaled dot-product cross-attention from the street branch as query to the BEV branch as key and value, including a learned relative positional bias $M$9. Fourth, Dual Dynamic Fusion adaptively merges the original street features $n$0 and geometry-enhanced features $n$1.

The DDF step is defined by channel-wise gates. A fusion weight vector $n$2 is computed by global average pooling and a linear layer $n$3: $n$4 The weighted components $n$5 are projected back to $n$6 channels by a $n$7 convolution, giving $n$8. An adaptive scaling mask is then formed as

$n$9

where $f:M\to M$0 is a $f:M\to M$1 convolution. The final fused feature is produced from $f:M\to M$2, optionally with $f:M\to M$3 as a residual, and the summary formulation states

$f:M\to M$4

Implementation details are given in concrete terms. All depthwise and pointwise convolutions use kernel size $f:M\to M$5 for proj and OSR, or $f:M\to M$6 in the gating mask. The attention uses typically 8 heads with $f:M\to M$7. The relative position bias $f:M\to M$8 is implemented as a learnable lookup table along the flattened $f:M\to M$9 grid. Dropout $x_i:\mathbb R\to\mathcal O_i$0 is applied to attention weights. GT Fusion is placed immediately after the backbone’s final feature map, at spatial resolution $x_i:\mathbb R\to\mathcal O_i$1 for $x_i:\mathbb R\to\mathcal O_i$2 input. Its street-view input is taken after the Causal Feature Extractor, $x_i:\mathbb R\to\mathcal O_i$3, to encourage the fused representation to focus on structural cues rather than low- or high-frequency style noise (Ouyang et al., 13 Mar 2026).

The empirical impact is reported through ablations and robustness evaluation. On CVACT$x_i:\mathbb R\to\mathcal O_i$4, EP-BEV has recall@1 of $x_i:\mathbb R\to\mathcal O_i$5, CLGT without GT Fusion has $x_i:\mathbb R\to\mathcal O_i$6, and full CLGT with GT Fusion has $x_i:\mathbb R\to\mathcal O_i$7. On CVUSA-C-ALL, EP-BEV achieves recall@1 of $x_i:\mathbb R\to\mathcal O_i$8, while CLGT with GT Fusion attains $x_i:\mathbb R\to\mathcal O_i$9. The qualitative description in attention-map visualizations is that GT Fusion helps the model maintain focus on road intersections and lane markings even under heavy snow, whereas the baseline’s attention drifts to sky or background textures (Ouyang et al., 13 Mar 2026).

4. GT Fusion as a handle-slide in 3d theories $i=1,\dots,m$0

In "Fusions and Dualities for 3d Theories $i=1,\dots,m$1," GT Fusion is not a feature-level or metric-level operation, but a topological move in a three-manifold together with its field-theoretic avatar. The paper defines the basic fusion operation as a special handle-slide, specifically a connected sum of two matter circles, or unknots, in $i=1,\dots,m$2. On the level of vortex or quiver partition functions, it is encoded by a “fusion identity” that expresses the contribution of a fused matter circle of parameter $i=1,\dots,m$3 as a sum over an intermediate flux $i=1,\dots,m$4 (Cheng, 2024).

The topological dictionary is explicit. Matter circles are unknotted loops $i=1,\dots,m$5 carrying OV-defect data, and each contributes one chiral multiplet. Gauge circles are surgery components, framed links along which one performs Dehn surgery to build $i=1,\dots,m$6, and each gauge circle $i=1,\dots,m$7 gives a $i=1,\dots,m$8 gauge factor. The linking number between a matter circle and a gauge circle is the matter’s charge. A standard Kirby handle-slide of one surgery circle over another realizes a change of basis in the gauge CS-level matrix $i=1,\dots,m$9. By contrast, sliding a matter circle over another matter circle produces the connected sum of the two unknots. This special handle-slide is GT Fusion.

On the plumbing graph, the move is written

$M_3$00

and the paper states that the vortex partition functions obey exactly the corresponding fusion identity. In gauge-theory language, $M_3$01 is the contribution of a chiral multiplet of mass $M_3$02 and flux $M_3$03, while the right-hand side describes summing over an intermediate flux $M_3$04. The operation therefore identifies a geometric connected sum with an algebraic summation formula.

The paper further states that fusion can be combined with gauging, flips, and decoupling to produce all known abelian 3d dualities, including SQED–XYZ, S-duality, and ST-duality. The example singled out in the data is the SQED–XYZ cubic duality. One begins with two free chirals of masses $M_3$05 and $M_3$06, performs GT Fusion to obtain one circle of mass $M_3$07 plus the intermediate sum over $M_3$08, and then interprets the resulting plumbing graph as a 2–3 move taking two chiral nodes to three nodes meeting at one gauge node $M_3$09. The paper states that this exactly reproduces the cubic superpotential $M_3$10 in the dual, with the three meeting lines encoding the three chiral fields in the XYZ model (Cheng, 2024).

Gaugings and flips also receive topological realizations. Gauging a flavor $M_3$11 is implemented by linking the corresponding matter circle with a new gauge circle of framing $M_3$12, while flipping a mass or a charge corresponds to reversing the orientation of the matter circle or of $M_3$13. The paper’s coherent picture summarizes GT Fusion as the connected-sum handle-slide of two Ooguri-Vafa matter-circles in $M_3$14, yielding a single matter-circle whose mass parameter is the product of the two original masses, and presents it as a fundamental building block for the web of 3d $M_3$15 abelian dualities.

5. Topology-change-aware volumetric fusion and non-manifold grids

A related but distinct use of the phrase “geometric topology fusion” appears in "Topology-Change-Aware Volumetric Fusion for Dynamic Scene Reconstruction." The paper re-designs the classical TSDF-based fusion framework so that dynamic scenes with topology changes can be reconstructed by introducing a Non-manifold Volumetric Grid into both TSDF and Embedded Deformation Graph representations. The stated goal is to allow connectivity updates by cell splitting and replication, thereby resolving the conflict between fixed-connectivity surfaces and discontinuous deformation fields under topology changes (Li et al., 2020).

The paper begins from the standard volumetric fusion pipeline, in which a dense 3D grid of voxels stores a truncated signed distance $M_3$16 and a running weight $M_3$17, and new observations are fused via weighted averaging after warping the canonical volume into the live frame. In the classical setting, a surface mesh is extracted from zero-crossings of $M_3$18, and an Embedded Deformation Graph models non-rigid motion. Both the mesh and the EDG are fixed-connectivity topological models. The paper identifies two failures under real topological change: the regularizer on EDG cannot break edges and therefore leaves “ghost” links across newly separated parts, and the mesh cannot naturally split or merge, producing blending artifacts or self-intersections.

The replacement is a non-manifold volumetric cell structure. The EDG grid is a regular lattice of EDG cells $M_3$19, each with exactly 8 corner nodes $M_3$20. The TSDF grid resides inside EDG cells; if the EDG refinement factor is $M_3$21, then each EDG cell contains $M_3$22 TSDF cells $M_3$23 and $M_3$24 TSDF voxels $M_3$25. Topology changes are first detected in registration as a set of cutting edges $M_3$26. For every EDG cell containing at least one such edge, the algorithm removes cutting edges from the internal node-graph, computes the connected components of the remaining 1-skeleton, duplicates the cell into one new cell per connected component, creates virtual nodes as needed so each duplicate again has 8 corners, and restores inter-cell adjacency under specified matching rules. TSDF splitting is analogous, with real-node voxels inheriting pre-existing $M_3$27 and $M_3$28, virtual voxels adjacent to a real negative-$M_3$29 node receiving $M_3$30, and remaining virtual voxels set to $M_3$31.

The full pipeline alternates deformation registration and topology fusion. Registration includes a rigid step by projective ICP and a non-rigid step that solves for local rotations, translations, and line-process weights $M_3$32 on adjacent EDG node-edges by minimizing

$M_3$33

Edges with $M_3$34 are marked in forward registration, a lightweight backward registration is performed, and the intersection produces a robust $M_3$35. Topology fusion then calls the cell-splitting procedure, propagates the split to embedded TSDF cells, warps each TSDF voxel by the updated EDG-driven deformation field, and fuses a new observation by

$M_3$36

Mesh extraction is handled by an extended marching cubes over non-manifold TSDF cells (Li et al., 2020).

The paper reports both complexity and evaluation. In regions without topology events, the overhead per frame is identical to a classical fusion method: $M_3$37 for warp and fuse and $M_3$38 for registration. For topology events, candidate EDG cells undergo only constant-size local graph operations, so the amortized per-frame overhead remains approximately $M_3$39 with a small constant factor. On a synthetic dataset with single-cut, multiple non-intersecting cuts, and intersecting cuts, the method achieves $M_3$40 mean point-to-surface distance post-cut versus $M_3$41 for VolumeDeform and $M_3$42–$M_3$43 for template-based methods, and reduces off-surface vertices by $M_3$44–$M_3$45. The limitations stated in the paper are that splitting events are handled robustly, whereas true merging, such as genus increase when two parts come into contact, would require merging EDG nodes and cells and is future work; the CPU-based prototype runs at about $M_3$46 FPS (Li et al., 2020).

6. Comparative interpretation, guarantees, and recurrent misconceptions

Across these works, “fusion” refers to materially different operations. In the multi-sensor time-series setting, one fuses delay-coordinate reconstructions from several observation functions and then corrects local linear dependencies. In CLGT, one fuses street-view and BEV feature maps so that explicit road topology can regularize cross-view correspondence. In 3d theories $M_3$47, one fuses two matter circles through a connected sum whose algebraic image is a partition-function identity. In dynamic scene reconstruction, one fuses depth into a canonical TSDF while allowing topological change through non-manifold connectivity updates (Solomon et al., 2020).

The guarantee structure is equally heterogeneous. The time-series framework explicitly invokes Takens’ theorem and its multivariate extension to justify topological embedding of the reconstructed orbit when $M_3$48 is large enough; however, it also states that no additional convergence proof is provided for the geometry. The CLGT formulation provides empirical evidence through recall@1 and attention-map behavior, not a topological embedding theorem. The 3d-theory formulation is exact at the level of the stated fusion identity and plumbing-graph move. The volumetric reconstruction framework provides algorithmic correctness for cell splitting and replication, together with quantitative post-cut error reductions, but states that true merging is future work (Ouyang et al., 13 Mar 2026).

A common misconception would be to treat GT Fusion as naming one fixed mathematical formalism. The literature instead assigns the label to at least four different constructions: a joint-delay-embedding metric, a BEV-guided cross-attention module, a handle-slide identity in $M_3$49, and a topology-change-aware volumetric-fusion redesign. Another misconception would be to assume that “topology” always means the same object. In these works it can mean manifold reconstruction and persistence features, road-network layout, connected sums of framed unknots, or connectivity changes in a canonical volume. This suggests that GT Fusion is best understood as a family resemblance term whose meaning is determined by domain, representation, and the specific fusion target rather than by a single canonical definition (Cheng, 2024).

The broader significance of the term therefore lies in the repeated attempt to make fusion sensitive to structure that a naive metric or concatenation can ignore. In one setting that structure is local tangency among sensor observation functions; in another it is stable road-intersection and junction geometry across viewpoint changes; in another it is the surgical calculus of matter and gauge circles; and in another it is the possibility of discontinuous deformation and topological splitting. The literature consistently uses GT Fusion to name fusion procedures that are meant to be structurally informed rather than purely aggregative (Li et al., 2020).