ChartPointFlow: Topology-Aware 3D Flow Model

• ChartPointFlow is a topology-aware generative model for 3D point clouds that uses invertible neural charts to map Euclidean latent spaces onto complex surfaces.
• It integrates permutation-invariant encoding, mixture-of-flows generation, and unsupervised chart assignment to accurately capture holes, intersections, and disconnected components.
• Empirical results on benchmarks like ShapeNet demonstrate superior performance in reconstruction (EMD) and unsupervised part segmentation compared to state-of-the-art models.

ChartPointFlow is a topology-aware, flow-based generative model for 3D point cloud surfaces. It represents each object as a union of invertible neural “charts,” each mapping a Euclidean latent space to a continuous subset of the 3D surface. By assigning chart labels to points in an unsupervised manner, ChartPointFlow preserves and models complex topological features—including holes, intersections, and disconnected components—while supporting semantic decomposition into subparts. The framework combines permutation-invariant shape encoding, mixture-of-flows point generation, and explicit chart regularization, setting new standards in generative quality and part segmentation for point cloud data (Kimura et al., 2020).

1. Model Architecture and Generative Framework

ChartPointFlow consists of four integrated neural components: a permutation-invariant feature encoder, a prior flow, a chart predictor, and a point generator flow. The feature encoder $E(X)$, implemented in a PointNet style, processes a point cloud $X = \{x_1, ..., x_M\}$ to produce a Gaussian posterior $q_E(s|X)$ over the global shape code $s \in \mathbb{R}^d$. The prior flow $G$ transforms $s$ into a base latent variable $w$, with a tractable prior $p_G(s)$ induced via normalizing flows. The chart predictor $C(x, s)$ yields posterior probabilities $\pi_C(y|x, s)$ over $X = \{x_1, ..., x_M\}$0 chart labels $X = \{x_1, ..., x_M\}$1 for each point, approximated through Gumbel–Softmax relaxation for differentiable sampling. The point generator $X = \{x_1, ..., x_M\}$2, implemented as a Glow-style flow, maps latent variables $X = \{x_1, ..., x_M\}$3 conditioned on $X = \{x_1, ..., x_M\}$4 and $X = \{x_1, ..., x_M\}$5 into 3D space.

Given $X = \{x_1, ..., x_M\}$6 and $X = \{x_1, ..., x_M\}$7, the generative process samples, for each point, $X = \{x_1, ..., x_M\}$8 from an auxiliary network and $X = \{x_1, ..., x_M\}$9, then outputs $q_E(s|X)$0. Under an i.i.d. assumption across points, the marginal likelihood of $q_E(s|X)$1 factorizes as

$q_E(s|X)$2

with $q_E(s|X)$3 given by the flow’s change-of-variable formula.

2. Chart-Based Topology Modeling

Each chart $q_E(s|X)$4 is an invertible diffeomorphism $q_E(s|X)$5, realized by restricting $q_E(s|X)$6 to the chart label. The union of chart images $q_E(s|X)$7 reconstructs the full 3D surface. By mapping from a ball-like latent space only over topologically trivial regions, each chart can be learned by a continuous flow network. The combination of several such maps allows ChartPointFlow to model surfaces with arbitrary genus and disconnected components, circumventing topological constraints that afflict single-flow models. The chart assignment mechanism, learned unsupervised, partitions the shape into continuous regions with sharp boundaries analogous to charts in differential geometry.

3. Latent Variable Structure and Unsupervised Label Assignment

In ChartPointFlow, each chart’s latent variable $q_E(s|X)$8 and label $q_E(s|X)$9 are governed by simple priors $s \in \mathbb{R}^d$0 and $s \in \mathbb{R}^d$1. During training, the chart posterior $s \in \mathbb{R}^d$2 is inferred for each point by the chart predictor network, with Gumbel–Softmax used to sample a one-hot surrogate for differentiability. This approach achieves an unsupervised, data-adaptive partitioning of the object's surface among the $s \in \mathbb{R}^d$3 charts without requiring subpart annotations. Assignment regularization is imposed to ensure both near-exclusivity of chart usage per point and broad chart coverage across the object.

4. Topology Preservation via Mixture-of-Flows

A single continuous flow between Euclidean spaces can only preserve the Euler characteristic—and is therefore unable to generate tori, holes, or disconnected shapes from a simple latent ball. ChartPointFlow addresses this limitation through a mixture of invertible flows, each “stitching” a locally flat latent domain into a region of the surface. The union of chart images produces the desired global topology, supporting nontrivial genus and manifold structure, while each local chart avoids the tearing or gluing required by single-flow architectures. This enables ChartPointFlow to faithfully recreate multimodal, non-simply-connected geometric surfaces.

5. Objective Function and Optimization Strategy

Model fitting proceeds by maximizing a joint evidence lower bound (ELBO) over both the shape code $s \in \mathbb{R}^d$4 and chart-labels $s \in \mathbb{R}^d$5. For $s \in \mathbb{R}^d$6 and $s \in \mathbb{R}^d$7, the ELBO is

$s \in \mathbb{R}^d$8

Expanding the flow likelihood and reparameterizing produces a tractable surrogate. Unregularized, the $s \in \mathbb{R}^d$9 might degenerate, causing chart overlap. To prevent this, a mutual-information term $G$0 is added:

$G$1

where $G$2 denotes entropy. Maximization encourages both uniform chart utilization and exclusivity in assignment. All modules $G$3 are optimized jointly using stochastic gradient ascent (Adam).

6. Empirical Performance and Segmentation Outcomes

ChartPointFlow demonstrates state-of-the-art results across multiple 3D surface benchmarks (Kimura et al., 2020). On synthetic 2D toy problems, ChartPointFlow accurately recovers holes and disconnected regions with as few as two charts, outperforming single-flow models such as Glow and FFJORD, which exhibit discontinuities and artifacts. On ShapeNet Core.v2 (classes: airplane, chair, car), ChartPointFlow with 24–28 charts achieves superior Earth Mover's Distance (EMD) and 1-Nearest-Neighbor Accuracy (1-NNA): for airplane generation, 1-NNA is 65.1% (ideal 50%), outperforming SoftFlow at 69.4% and PointFlow at 75.1%. For reconstruction, ChartPointFlow attains EMD $G$4 (airplane) and $G$5 (chair), representing a 10–30% improvement over AtlasNet, AtlasNet V2, PointFlow, and SoftFlow.

Qualitative analysis shows that chart assignments correspond to semantic object subparts (e.g., airplane’s wings, fuselage, tail; chair’s legs, seat, backrest, armrests; car’s wheels, doors, body panels). In unsupervised part segmentation experiments on ShapeNetPart, ChartPointFlow attains NMI/purity of 0.30/0.80 (airplane) and 0.35/0.86 (chair), surpassing AtlasNet variants despite the absence of part labels during training.

Dataset / Metric	ChartPointFlow	Best Competing Method	Score (ChartPointFlow)
Airplane 1-NNA	SoftFlow	69.4%	65.1%
Chair EMD	AtlasNet V2	6.17e-2	4.62e-2
Segmentation purity	AtlasNet V2	0.68	0.80

7. Extensions, Applications, and Integration Potential

The chart-based mixture-of-flows paradigm underlying ChartPointFlow is adaptable to additional modalities beyond static 3D surfaces. A plausible implication is that the chart/decomposition framework can be combined with latent-structure priors developed for sequential or regime-shifting data, such as those from group-fused LASSO latent Gaussian mixture-of-experts models employed in change-point detection (Kei et al., 5 Mar 2026). By integrating a group-fused penalty for latent chart means, ChartPointFlow could support piecewise-constant regime modeling, enabling full model-based detection of structural transitions in point cloud sequences or spatiotemporal data. Extensions are feasible by substituting or augmenting the chart predictor and generator to accommodate heterogeneous covariates or multimodal sensory input, as suggested by integration guidelines for latent-space mixture models. This suggests significant capacity for ChartPointFlow as a fully model-based detection and segmentation engine across high-dimensional surface representations and sequentially indexed data streams.