TopGeoFormer: Topology & Geometry Recovery

• TopGeoFormer is an end-to-end architecture for point cloud recovery that couples down-preservation with cascaded up-preservation to maintain both topological and geometric embeddings.
• It introduces InterTwining Attention to effectively merge local structural features with global shape codes, ensuring fine structure details are retained during restoration.
• The framework leverages explicit geometry and topological loss functions to drive high-fidelity recovery, demonstrating superior performance across varied datasets.

Searching arXiv for TopGeoFormer and closely related names to ground the article. TopGeoFormer is an end-to-end architecture for point cloud recovery that is explicitly designed to preserve both topological and geometric embeddings throughout a sequential sampling-and-restoration pipeline. In the formulation of "Preserving Topological and Geometric Embeddings for Point Cloud Recovery" (Zhou et al., 25 Jul 2025), the model couples a Down-Preservation phase with two cascaded Up-Preservation phases, revisits local feature extraction through a continuous mapping of neighboring-point relations, introduces InterTwining Attention to merge topology and geometry, and optimizes recovery with a full geometry loss and a topological constraint loss. The result is a recovery-oriented framework in which sampling is not treated as a disposable preprocessing step, but as the first half of a structure-preserving transformation.

1. Problem setting and conceptual scope

TopGeoFormer addresses point cloud recovery, defined as a sequential process of sampling / down-preservation and restoration / up-preservation. This framing is narrower than generic reconstruction and differs from completion, because recovery assumes that the entire input point cloud is available initially, but must be reduced and later restored under communication or processing constraints. It also differs from ordinary upsampling, because recovery must preserve informative structure already at the downsampling stage (Zhou et al., 25 Jul 2025).

The task begins from a point set

$$P = \{p_i \mid i=1,\dots,N\} \in \mathbb{R}^{N \times 3},$$

and learns a continuous mapping $f$ that produces a coarse sampled cloud $R_0$ and then a restored cloud $R$. The paper emphasizes continuity as a structural principle: if the mapping $f$ is continuous and local Euclidean neighborhoods overlap sufficiently, topological and geometric properties can be preserved. In this setting, geometric embedding refers to sampled point coordinates and coarse surface coverage, whereas topological embedding refers to local neighborhood relations and manifold structure. This suggests that TopGeoFormer treats recovery as preserving a latent organization of the original space, not merely approximating point positions.

The motivation is a concrete limitation of established pipelines. Conventional sampling methods such as FPS preserve coverage but may destroy fine structure, while restoration methods often receive only coarse information and therefore recover overly smooth shapes. TopGeoFormer is organized around the premise that sampling quality and restoration quality are coupled: if the sampled representation fails to preserve local structure, the decoder cannot recover it later.

2. Topological and geometric representations

A central design choice is the construction of a topological embedding from local relative geometry. For each point $p_i$, TopGeoFormer computes local relational vectors

$$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$

where $\mathcal{N}_P(p_i)$ is the set of $k$-nearest neighbors of $p_i$. These subtraction vectors encode magnitude, direction, and local connectivity. They are then mapped by an MLP from

$f$0

to

$f$1

which the paper interprets as a continuous mapping into a higher-dimensional topological space. A max-pooling operator aggregates neighborhood features to make the descriptor more robust to local perturbations (Zhou et al., 25 Jul 2025).

The Down-Preservation (DP) phase uses PointNet++-style hierarchical abstraction with three set abstraction (SA) layers, FPS + kNN grouping, and produces a global shape code

$f$2

The first SA layer yields sampled points

$f$3

as the geometric embedding, together with local features

$f$4

The local feature extraction is written as

$f$5

with

$f$6

Local and global information are then fused by

$f$7

where $f$8 is the point-shape fusing representation.

This representational split is essential to the model’s identity. $f$9 carries coarse geometry, $R_0$0 carries transformed neighborhood relations, and $R_0$1 combines local point awareness with object-level context. A plausible implication is that the architecture avoids collapsing manifold structure into a purely global latent code, which is a common source of oversmoothing in coarse-to-fine reconstruction pipelines.

3. Architecture: Down-Preservation, Up-Preservation, and InterTwining Attention

The full pipeline consists of one DP phase followed by two cascaded UP phases. The model’s core fusion mechanism is InterTwining Attention (ITA), which is used in both phases to merge topological embeddings and geometric / shape embeddings (Zhou et al., 25 Jul 2025).

In the DP stage, ITA operates on the fused representation $R_0$2 together with a topological embedding computed on the downsampled points,

$R_0$3

The attention update is

$R_0$4

and

$R_0$5

where $R_0$6 is Softmax, $R_0$7 is an MLP with two linear layers and ReLU, and $R_0$8 denotes elementwise interaction / modulation. The attention output is transformed as

$R_0$9

and the sampled output becomes

$R$0

Conceptually, ITA is described as forming a learnable shape context built from three kinds of information: point-wise features, point-shape-wise features, and intra-shape features. In the DP phase, this allows the network to move sampled seeds while maintaining local manifold structure. The paper’s explanation is not that attention merely aggregates local cues, but that topology guides how the learned shape code is queried.

The Up-Preservation (UP) phase restores density progressively through two stages. Let $R$1 denote the current input to a UP stage, with $R$2 in the first stage. Because seed features are not transmitted directly, the model computes an upsampling feature $R$3 that acts as a locally weighted propagation of features using nearby points, integrating topological information into the restoration process. In parallel, $R$4 is encoded to obtain $R$5. Before the main up-preserving attention, ITA enhances the query through

$R$6

with $R$7 if not present. The model also computes $R$8 and corresponding neighbor indices.

The ensuing Up-Preserving Attention (UPA) uses $R$9 and $f$0 as 1D convolution layers applied to $f$1 and $f$2, an $f$3 ResMLP to form the value vector $f$4, and a further MLP $f$5 composed of a linear 2D layer, ReLU, and transposed 2D convolution to merge topological features, upsampling features, and intra-shape features. The predicted displacements refine the current point set, and after two UP phases the final restored cloud $f$6 is produced.

4. Optimization objectives and training protocol

TopGeoFormer uses two explicit supervisory terms: a full geometry loss and a topological constraint loss (Zhou et al., 25 Jul 2025). The geometry term departs from standard coarse-to-fine Chamfer supervision, where intermediate predictions are matched only to FPS-reduced targets. Instead, every intermediate prediction is matched to the full-resolution ground truth $f$7:

$f$8

where $f$9 is the DP output, $p_i$0 is the first UP output, and $p_i$1 is the final recovery. Chamfer Distance is given as

$p_i$2

The paper characterizes this as an intentionally “inconsistent” matching between coarse-to-fine generations and the original target, arguing that it forces each stage to align with the full shape and improves geometric detail reconstruction.

The topological term constrains embedding variance:

$p_i$3

with

• $p_i$4,
• $p_i$5,
• $p_i$6,
• $p_i$7.

The final objective is

$p_i$8

The loss weighting indicates that geometry is heavily prioritized, while the topological term acts as a regularizer for manifold consistency. The paper states that $p_i$9 gives a slight but consistent improvement, whereas $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$0 is a main driver of reconstruction quality.

Training is reported with PyTorch, a single RTX 3090, batch size 32, 120 epochs, the Adam optimizer, initial LR 0.005, learning rate halved every 30 epochs, and data augmentation by mirroring, scaling, rotation. These implementation details place TopGeoFormer within a relatively compact training regime rather than a large-scale pretraining setup.

5. Evaluation, ablations, and reported performance

The empirical study spans ModelNet40 and ScanObjectNN for object-level recovery, with three input settings—uniform set, random set, and partial set—and PU-147, KITTI-10, and ScanNet-30 for generalization and large-scene evaluation. The reported metrics are Chamfer Distance (CD), Hausdorff Distance (HD), and Earth Mover’s Distance (EMD) (Zhou et al., 25 Jul 2025).

On object-level recovery, the paper reports that TopGeoFormer achieves the best overall performance. For the random set, it reports CD 0.62, HD 0.29, and EMD 2.82. For the partial set, it reports CD 0.26, HD 0.18, and EMD 2.39. Relative to the second best method, the paper states CD reductions of 17% on uniform, 27% on random, and 32% on partial inputs. The recovery baselines include PU-GCN, Dis-PU, PointSE, RepKPU, SampleNet, APES, and PointSE$$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$1, with additional appendix comparisons to Grad-PU, SPU-IMR, SnowflakeNet, SeedFormer, and SVDFormer.

The paper also evaluates the sampling stage directly. On the random set at 2048 points, sampled CD is reported as 3.71 for SampleNet, 1.64 for FPS, 2.41 for APES, and 1.51 for TopGeoFormer. At 512 points, the reported values are 8.76, 6.42, 9.17, and 5.88, respectively. These numbers are used to support the claim that the sampling module is not only geometry-preserving but also recovery-oriented.

On larger scenes, TopGeoFormer outperforms PointSE$$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$2 on both KITTI-10 and ScanNet-30, with KITTI-10 CD 0.73 vs 1.39 and ScanNet-30 CD 1.26 vs 1.65. The qualitative interpretation reported in the paper is that the model yields better outline continuity and fewer geometric artifacts. It is also described as preserving fine structures such as propeller blades, wing edges, fuel tanks, leaves, chair boundaries, holes and openings, and other thin structures.

The ablation study isolates the contribution of major components. Reported findings include that using $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$3 together with $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$4 improves results over using $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$5 alone, ITA improves both sampling and recovery, the proposed UP module outperforms the SPD decoder, $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$6 gives a noticeable gain, and $$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$7 gives a smaller but consistent gain. The best performance comes from combining topological embeddings, geometric embeddings, ITA, UP, geometry loss, and topological loss.

Efficiency is also reported explicitly. The small model variant has 0.46M parameters and 4G FLOPs, the full model has 2.03M parameters and 14G FLOPs, and ITA itself has only 0.08M parameters. The paper further claims that the model trains faster than PointSE$$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$8 because it avoids repeated redundant grouping computations.

6. Scope, limitations, and relation to similarly named models

The paper identifies several limitations. TopGeoFormer is designed to favor non-uniform inputs, is trained only on complete objects, and is not suited for predicting largely missing regions (Zhou et al., 25 Jul 2025). Scene-level recovery generalization is presented as an open direction. These constraints matter when situating the model relative to completion-oriented methods: TopGeoFormer is not intended for settings where substantial regions are absent and must be hallucinated from partial evidence.

A frequent source of confusion is nomenclature. Despite its name, TopGeoFormer is not part of the family of geospatial transformers for environmental sensing, mobility modeling, urban forecasting, or text geocoding. It is unrelated to GeoFormer for NO$$\mathcal{D} = \{p_i - p_j \mid p_j \in \mathcal{N}_P(p_i)\},$$9 estimation from Sentinel-5P imagery (Khirwar et al., 2024), GeoFormer as a decoder-only GPT model for human mobility prediction (Solatorio, 2023), GeoTransformer for urban forecasting with geospatial attention (Jia et al., 2024), GeoNorm for toponym resolution (Zhang et al., 2023), GeoReFormer for 3D lane segment detection and topology reasoning (Abraham et al., 27 Mar 2026), and the HAKE-based geometry-aware GeoKG embedding model in "Geometric Feature Enhanced Knowledge Graph Embedding and Spatial Reasoning" (Hu et al., 2024). In those works, “Geo” and sometimes “topology” refer to geospatial relations, urban regions, or road-graph structure; in TopGeoFormer, they refer to the preservation of point-cloud manifold structure across sampling and restoration.

Within its own scope, TopGeoFormer’s significance lies in treating recovery as a coupled topology-and-geometry preservation problem. The architecture, losses, and ablations are all organized around that premise. A plausible implication is that the work’s main legacy is methodological rather than terminological: it argues that recovery quality depends not only on decoder expressivity, but on whether both the reduced representation and the restoration trajectory preserve local structure from the outset.