Source-linked AI summary
Hyper^2: Unleashing Hyperbolic Geometry's Full Potential via Dual-Space Consistency
Guantian Zheng, Haiyang Xu, Tianyu Gao
TL;DR
HyperbolicCD’s hyperbolic loss yields modest gains because it back-propagates through a Euclidean encoder, creating a cross-geometry mismatch. Hyper^2 aligns the encoder and loss with the same hyperbolic distance form and shared curvature, achieving super-additive completion gains with low overhead. The evidence supports geometric consistency as the relevant design principle for point cloud completion, while transfer to other architectures, curvatures, geometries, and tasks remains open.
Problem
HyperbolicCD reports modest 3–7% Chamfer reductions because its hyperbolic loss is paired with a Euclidean encoder, creating a cross-geometry mismatch.
Method
Hyper^2 applies arcosh(1+αd^2) both as a refinement-attention positional bias and as the hyperbolic Chamfer loss, using one shared α.
Results
−22.9% Chamfer on ShapeNet-55 and −37.5% on 21 unseen ShapeNet-34 categories are achieved, with approximately 1.6% additional FLOPs.
Takeaways & Limitations
The indicators rise from (0.68,39%) to (0.95,87%) only in the dual-space configuration, supporting geometric consistency across encoder and loss as the paper’s central conclusion.
Takeaways & Limitations
The encoder design is tied to an explicit incompleteness distance field, single global curvature is used, and transfer to other geometries or 3D tasks remains untested.
Abstract
from arXiv · showhide
HyperbolicCD pioneered hyperbolic geometry for point cloud completion by replacing the Euclidean Chamfer distance with arcosh(1+alpha||x-y||^2), but the reported gains are modest (3-7% Chamfer reduction across SeedFormer, PointAttN and PMP-Net backbones on PCN and ShapeNet-55). We argue the bottleneck lies elsewhere: the loss is hyperbolic but the encoder it back-propagates through is Euclidean, so the position-dependent supervision of the loss is averaged away by the chain rule before it reaches the parameters. We call this a cross-geometry mismatch, and make it testable through two model-agnostic indicators, feature-loss correlation r_FL and effective gradient utilisation u_G. On an SVDFormer backbone trained with HyperbolicCD's loss alone we measure (r_FL, u_G) = (0.68, 39%). We propose Hyper^2, a dual-space consistency framework that extends HyperbolicCD by reusing the identical arcosh(1+alpha d^2) functional form as a positional bias on the refinement attention (a hyperbolic distance encoding), paired with HyperbolicCD's hyperbolic Chamfer loss under a single shared curvature alpha. Both operators are O(N log N) scalar non-linearities on Euclidean distances and together add only ~1.6% FLOPs over SVDFormer. Hyper^2 delivers -22.9% Chamfer on ShapeNet-55 over SVDFormer (well above the 13.2% linear sum of the -12.0% loss-only and -1.2% encoding-only single-space ablations) and -37.5% on the 21 unseen ShapeNet-34 categories. The two indicators remain essentially flat for any single-space configuration but jump together to (0.95, 87%) only when both encoder and loss are hyperbolic, supporting the claim that geometric consistency across encoder and loss, rather than either operator alone, is what enables hyperbolic supervision in point cloud completion. Code is available at https://github.com/Ethan-Zheng136/Hyper-2.
1 Introduction
Point cloud completion must recover hierarchical 3D structure, but Euclidean Chamfer treats coarse and fine errors similarly. Hyper^2 addresses the cross-geometry mismatch by aligning hyperbolic encoding and loss, producing super-additive gains with low overhead.
- Point cloud completion recovers complete 3D shapes from partial observations for autonomous driving, robotic manipulation, and augmented reality.
- Euclidean Chamfer cannot distinguish qualitatively different coarse and fine geometric errors because it treats them proportionally to point-wise distances.
- HyperbolicCD reports only 3–7% Chamfer reduction across SeedFormer, PointAttN, and PMP-Net on PCN and ShapeNet-55.
- Hyper^2 applies the same arcosh(1+αd^2) form as a refinement-attention positional bias and a hyperbolic Chamfer loss with one shared curvature α.
- −22.9% Chamfer on ShapeNet-55 exceeds the 13.2% linear sum of the −12.0% loss-only and −1.2% encoding-only ablations.
2 Related Work
Prior completion methods largely use Euclidean representations, while hyperbolic learning has mainly addressed hierarchical embeddings and other non-Euclidean settings. Hyper^2 targets geometric consistency between feature and loss spaces in point cloud completion.
- Early learning-based methods: Early completion methods established coarse-to-fine generation but struggled with complex geometric details because of limited capacity.
- Transformer-based methods: Transformer-based methods capture global structure effectively but operate entirely in Euclidean space and cannot explicitly encode hierarchical relationships.
- Refinement strategies: Refinement methods support iterative detail generation but apply unified strategies across regions, limiting adaptability to different missing patterns.
- Cross-modal methods: SVDFormer avoids paired RGB-point cloud data by using self-projected depth maps and an incompleteness encoding for coarse-to-fine refinement.
- Foundational theory: Hyperbolic space represents hierarchical data through exponential volume growth, motivating its use for structured geometric relationships.
- Hyperbolic neural networks: Hyperbolic neural networks and graph methods use tangent-space operations to provide non-Euclidean representations while maintaining computational efficiency.
- Vision applications: Vision-side hyperbolic methods focus on embedding-space design without considering consistency with downstream loss functions.
- HyperbolicCD: HyperbolicCD introduced arcosh(1+α∥x−y∥^2) as a point-cloud completion loss, retaining O(N logN) KNN complexity but leaving the encoder Euclidean.
3 Hyper2 Framework
Hyper^2 extends hyperbolic Chamfer matching into the refinement encoder, using the same arcosh(1+αd^2) form and shared curvature α to align encoder and loss geometry. The framework preserves efficient Euclidean-distance computation while supplying sub-linear positional bias and hyperbolic supervision throughout coarse-to-fine completion.
- Hyperbolic Chamfer Distance: Hyperbolic Chamfer Distance replaces Euclidean nearest-neighbour distances with arcosh(1+α∥x−y∥^2), retaining KNN-based computation without full Poincaré-ball projection.With constant α, the formulation avoids boundary instability and retains O(N log N) KNN complexity.
- Hyperbolic Incompleteness Encoding: Hyper^2 replaces SVDFormer’s Euclidean incompleteness encoding with arcosh(1+αd^2) as a positional bias on refinement self-attention.The change preserves the surrounding feature embedding, self-attention, cross-attention, decoder, and feature pathway.
- Shared Curvature and Supervision: The three stages P0, P1, and P2 are each supervised by the hyperbolic Chamfer loss, while encoder and loss share one scalar curvature α.This implements end-to-end geometric consistency across the refinement pipeline.
- Hyperbolic Incompleteness Encoding: Near observed regions, the encoding is approximately linear, while for distant points it becomes log-compressed, preventing distance alone from monopolising attention.This encourages far points to be disambiguated through feature context and coarse global structure.
- Hyperbolic Chamfer Distance: The arcosh transform is sub-linear rather than exponential: its value is approximately 2αd near zero and log-compressed for large d, with a bounded gradient.The resulting saturation gives coarse and fine errors comparable priority instead of allowing far outliers to dominate.
- Position-Dependent Supervision: The loss gradient and encoder positional bias carry the same arcosh′(1+αd^2) position-dependence, allowing that dependence to survive the chain rule.The intermediate attention block is linear in its inputs; under Euclidean encoding, attention softmax instead averages the loss’s position-dependence across points.
4 Experiments
Experiments show that Hyper^2 improves completion across ShapeNet, PCN, and KITTI while adding minimal computational overhead. Ablations indicate that the strongest gains and diagnostic improvements arise specifically from jointly hyperbolic encoding and loss.
- ShapeNet-55/34: Hyper2 reduces ShapeNet-55 CD-Avg from 0.83 to 0.64 (−22.9%) and improves F1 from 0.451 to 0.523.The relative CD reductions are −22.9%, −21.4%, and −22.3% across simple, moderate, and hard difficulty levels.
- ShapeNet-55/34: Hyper2 reduces ShapeNet-34 CD-Avg by 17.3% on seen categories and by 37.5% on 21 unseen categories.The reported values are 0.75→0.62 for seen categories and 1.28→0.80 for unseen categories.
- PCN: Hyper2 reduces PCN CD-Avg from 6.54 to 6.36 (−2.8%) over SVDFormer and improves F1 from 0.841 to 0.854.It is best on every category, with DCD falling from 0.536 to 0.528.
- KITTI: Hyper2 achieves KITTI Fidelity 0.026 and MMD 0.109, improving over the SVDFormer reproduction’s 0.052 and 0.145.Fidelity checks visible-region preservation, while MMD probes hallucinated unobserved structure.
- Ablation Studies: Hyperbolic loss alone gives −12.0% CD, encoding alone gives −1.2%, and both together give −22.9%, exceeding the 13.2% linear-sum prediction.The dual-space configuration also raises rFL and uG together to (0.95,87%), whereas single-space configurations remain near baseline.
- Efficiency: Hyper2 increases FLOPs from 12.3G to 12.5G (+1.6%), keeps parameters at 23.1M, and increases inference time from 47ms to 48ms (+2.1%).The two operators are scalar non-linearities on Euclidean distances with O(N logN) cost.
5 Conclusion
Hyper2 aligns hyperbolic distance encoding and hyperbolic Chamfer supervision under shared curvature, addressing the cross-geometry mismatch. It improves benchmark performance and indicator alignment, while remaining constrained by architectural and evaluation limitations.
- 22.9% Chamfer reduction on ShapeNet-55 and 37.5% on unseen ShapeNet-34 accompany a more modest 2.8% reduction on PCN.Hyper2 adds approximately 1.6% FLOPs.
- Hyper2 places the same arcosh(1 + αd^2) form in refinement attention encoding and the Chamfer-style training loss.Both uses are bound by a shared curvature α.
- The indicators rise from (0.68, 39%) to (0.95, 87%) after closing the mismatch between Euclidean encoding and hyperbolic loss.The indicators are computed from one validation-batch forward-and-backward pass.
- Transfer beyond SVDFormer-style explicit incompleteness fields, adaptive curvature, causal validation, and other manifolds or 3D tasks remain untested.The paper identifies these as limitations or future directions.