Technical Summary: Hyper2: Unleashing Hyperbolic Geometry's Full Potential via Dual-Space Consistency

Problem Statement

Point cloud completion aims to recover complete 3D shapes from partial observations, a task critical for autonomous driving and robotics. A fundamental challenge is that 3D shapes possess hierarchical geometric structures where coarse errors (e.g., a missing wing) differ qualitatively from fine errors (e.g., surface roughness). Standard Euclidean Chamfer Distance (CD) treats these errors proportionally to point-wise distances, failing to distinguish between them.

While HyperbolicCD [1] introduced hyperbolic geometry to this domain by replacing the Euclidean CD loss with an $\text{arcosh}(1+\alpha\|x-y\|^2)$ formulation, the reported performance gains were modest (3–7% CD reduction). The authors argue that the bottleneck lies in a cross-geometry mismatch: the loss function operates in hyperbolic space, but the encoder and refinement modules of modern backbones (e.g., SVDFormer, SeedFormer) remain Euclidean. Consequently, the position-dependent supervision provided by the hyperbolic loss is averaged out by the chain rule as it back-propagates through the Euclidean encoder, diluting the hierarchical signal before it reaches the network parameters.

Methodology: The Hyper2 Framework

The authors propose Hyper2, a dual-space consistency framework designed to align the geometric properties of the encoder and the loss function. The core insight is that geometric consistency across the entire network is required to preserve the benefits of hyperbolic supervision.

1. Dual-Space Consistency Principle

Hyper2 extends the functional form of the HyperbolicCD loss, $\text{arcosh}(1+\alpha d^2)$, from the loss side to the encoder side. Both components share a single scalar curvature hyperparameter $\alpha$, ensuring end-to-end geometric alignment.

2. Hyperbolic Incompleteness Encoding (Encoder Side)

Building upon the SVDFormer architecture [12], Hyper2 replaces the standard Euclidean incompleteness encoding with a Hyperbolic Distance Encoding.

• Mechanism: For each point $x_i$, the distance to the observed partial input $P_{in}$, denoted $d_i = \min_{y \in P_{in}} \|x_i - y\|$, is transformed via $h^H_i = \text{Sinusoidal}(\text{arcosh}(1 + \alpha d_i^2))$.
• Function: This scalar non-linearity is injected as a positional bias into the self-attention mechanism of the refinement module.
• Behavior: The $\text{arcosh}$ transform is sub-linear. For points near the observed region ($\alpha d^2 \ll 1$), the encoding is approximately Euclidean ($\sqrt{2\alpha}d$), preserving fine-grained details. For points far away ($\alpha d^2 \gg 1$), it saturates to a logarithmic compression ($\log(2\alpha d^2)$), preventing distant outliers from dominating the attention budget and forcing the model to rely on global structural context.
• Efficiency: This operation is a scalar non-linearity applied to Euclidean distances computed via standard KNN, maintaining $O(N \log N)$ complexity.

3. Hyperbolic Chamfer Distance (Loss Side)

The framework adopts the HyperbolicCD loss verbatim:
$$L_{\text{Hyper2}}(P, Q) = \frac{1}{|P|} \sum_{p \in P} \text{arcosh}(1 + \alpha \|p - \text{NN}(p, Q)\|^2) + \text{sym.}$$
Crucially, the curvature $\alpha$ is shared between the encoder encoding and the loss function. This shared parameter ensures that the gradient flow preserves the $\text{arcosh}$ shape end-to-end, implementing a "coarse-first, fine-later" training regime.

4. Diagnostic Indicators

To validate the claim of cross-geometry mismatch, the authors define two model-agnostic indicators:

• Feature-Loss Correlation ($r_{FL}$): The Pearson correlation between the encoder's incompleteness encoding scalar and the loss gradient magnitude.
• Effective Gradient Utilisation ($u_G$): The $R^2$ score of fitting the observed gradient magnitude to the theoretical hyperbolic gradient shape.
  The authors hypothesize that these indicators will remain low for single-space configurations (Euclidean encoder + Hyperbolic loss) but jump significantly only when both encoder and loss are hyperbolic.

Key Contributions

1. Diagnostic Indicators: The identification of cross-geometry mismatch as the primary bottleneck in prior hyperbolic completion methods. The authors define $r_{FL}$ and $u_G$ to auditable measure the consistency between feature encoding and loss supervision.
2. Hyper2 Framework: A dual-space consistency approach that reuses the $\text{arcosh}(1+\alpha d^2)$ form as a positional bias in the encoder and pairs it with a shared-curvature hyperbolic loss.
3. Empirical Evidence of Super-Additivity: Demonstration that the combined performance of dual-space consistency exceeds the linear sum of single-space improvements, validating that geometric alignment, rather than individual operators, drives the gains.

Experimental Results

Experiments were conducted on ShapeNet-55, ShapeNet-34 (including 21 unseen categories), PCN, and KITTI LiDAR scans.

• ShapeNet-55: Hyper2 achieves a 22.9% reduction in average Chamfer Distance (CD) compared to the SVDFormer baseline (0.83 $\to$ 0.64). This significantly outperforms the linear sum of single-space ablations (Hyperbolic loss only: -12.0%; Hyperbolic encoding only: -1.2%).
• ShapeNet-34 (Unseen Categories): The framework shows a 37.5% CD reduction on 21 unseen categories, suggesting superior generalization to complex hierarchical structures.
• PCN Dataset: A more modest 2.8% CD reduction over SVDFormer, consistent with the framework's utility being most pronounced where hierarchy recovery is most difficult.
• KITTI: On real-world LiDAR scans, Hyper2 achieves a Fidelity of 0.026 and MMD of 0.109, outperforming SVDFormer by roughly 2x in fidelity and 25% in MMD.
• Efficiency: The method adds only ~1.6% FLOPs and ~2.1% inference time over the baseline, as both operators are scalar non-linearities on Euclidean distances.

Diagnostic Validation

The ablation studies confirm the theoretical claims:

• Euclidean + Euclidean: $r_{FL} = 0.71, u_G = 44\%$.
• Euclidean + Hyperbolic (HyperbolicCD): $r_{FL} = 0.68, u_G = 39\%$ (performance drops slightly, indicating the mismatch dilutes the signal).
• Hyperbolic + Euclidean: $r_{FL} = 0.74, u_G = 46\%$.
• Hyperbolic + Hyperbolic (Hyper2): $r_{FL} = 0.95, u_G = 87\%$.
  The indicators only jump to high values when both ends of the network are hyperbolic, supporting the claim that geometric consistency is the driver of performance.

Significance and Claims

The paper claims that geometric consistency across the encoder and loss is the prerequisite for effective hyperbolic supervision in point cloud completion. It argues that a non-Euclidean loss is only as useful as the encoder it back-propagates through; without alignment, the position-dependent gradients are averaged away.

Hyper2 demonstrates that by extending the hyperbolic functional form to the encoder's positional bias and sharing the curvature parameter, one can unlock the full potential of hyperbolic geometry for hierarchical data. The authors position $r_{FL}$ and $u_G$ as cheap, model-agnostic tools for auditing future non-Euclidean methods.

Limitations noted by the authors:

• The encoder operator relies on an explicit incompleteness distance field (as in SVDFormer); adapting it to backbones without such fields (e.g., PoinTr) requires further architectural design.
• The framework uses a single global curvature $\alpha$; per-instance or per-region curvature is unexplored.
• The diagnostic indicators are correlational; while they jump together in the dual-space configuration, they do not strictly prove causation.
• Transfer to other manifolds (spherical, product) or 3D tasks (segmentation, scene flow) remains untested.