Real 3D sensors do not deliver clean shapes. Scans arrive jittered, sparse, and missing whole regions that sit in blind spots. Two kinds of prior knowledge could repair this damage. The first is topology: a chair is one connected piece with a known number of holes. The second is symmetry: the occluded left armrest should mirror the visible right one. Methods that use topology already exist. Methods that use symmetry run into two problems. Exact symmetry detection is expensive, and on noisy data it tends to find nothing, because noise breaks every exact symmetry. This note proposes looking for symmetry on a small, noise-filtered topological skeleton of the shape, meaning a graph of its persistent junctions, branches and loops, instead of on the raw points. The symmetries accepted on that skeleton then drive a regularizer that fills occluded regions and keeps the reconstruction's topology consistent.
In two sentences: first reduce a noisy scan to the handful of structural features that survive noise, and search for mirror or rotational symmetries there, where the search is small and stable. Then penalize the reconstructed surface when it breaks those symmetries or the skeleton's connectivity, so occluded parts are recovered from their visible counterparts rather than left as holes or floating fragments.
Core Claim
Our central claim is about cost and robustness together. On noisy, partially occluded point clouds, symmetry detection on a persistence-filtered skeleton graph should do two things. It should cost a small fraction of the downstream reconstruction time, unlike raw-point or full-automorphism detection. It should also recover correct symmetries at noise and occlusion levels where raw-point detection fails. Using the detected symmetries as a soft invariance and topology penalty should then reduce topological errors beyond what topology losses alone achieve: spurious components, missing loops, and erroneous handles. The new part is the order of operations. Topology is used first to make symmetry detection tractable and robust, and the resulting symmetry is then used to regularize topology. Prior work treats these as separate concerns.
Where This Sits
Symmetry detection overhead is well documented outside vision. The orbit problem is as hard as set-stabilizer-in-a-coset problems (Clarke et al. 1998). Finding canonical representatives is NP-hard in planning (Pochter et al.). Automorphism search can exceed solve time in mixed-integer programming and games (Hojny; Tewolde et al.). In lifted inference, preprocessing can eclipse inference (Singla et al.). These papers motivate our concern that detection must be budgeted, but all of them assume clean discrete inputs.
Two symmetry papers supply ingredients we reuse. Achieving Approximate Symmetry Is Exponentially Easier than Exact Symmetry shows that averaging over $O(\log |G|)$ random group elements captures most of the benefit that exact averaging over a group $G$ provides; 32 samples sufficed for $|G|=2^{20}$. It assumes the group is known in advance. We address the step it leaves open, which is finding a finite candidate group from damaged data. Natural Graph Networks shows that restricting symmetry reasoning to local structure keeps cost linear in edges, which supports reasoning on a sparse graph rather than a dense cloud. That work assumes uncorrupted graphs.
On the topology side, Guaranteed Simply Connected Mesh Reconstruction removes handles algebraically and prunes outliers, reaching one connected component where CrossSDF and Screened Poisson produce dozens. It constrains our design in two ways: it forces genus zero, and it treats inlier points as exact vertices, so it is exposed to jitter. CrossSDF fits a per-scene neural signed distance field (a network whose output is the distance to the surface, with sign indicating inside or outside) and keeps thin structures connected. It uses no symmetry. clDice and the linear-time supervoxel critical-component loss give differentiable or cheap connectivity penalties on voxel grids. We reuse the critical-component idea for our topology term.
SinPoint preserves topology during augmentation with bounded sine warps. We borrow it to build the continuous-deformation part of our corruption model. On sensing, Towards Autonomous Grading In The Real World documents how rule-based pipelines that were near-perfect in clean simulation fall to 14.3% success under real sensor noise. ESC handles permutation symmetry under tracking noise but not spatial symmetry. Neither work detects or exploits geometric symmetry.
Research Program
The pipeline has three stages.
The first stage builds a skeleton. From a noisy point cloud we fit a coarse occupancy field on a downsampled grid. We compute persistent homology once, not per training step. Persistent homology tracks which connected components and loops persist as a threshold is swept; short-lived features are treated as noise. The features whose persistence exceeds a noise threshold $\tau$ are kept, and $\tau$ is calibrated from the sensor's jitter scale. Junctions, endpoints and loop cycles become nodes, and branches become edges, each carrying length, thickness and relative position. A typical object yields tens of nodes where the scan has tens of thousands of points.
The second stage detects symmetry. Each pair of skeleton nodes with compatible attributes proposes a reflection plane. Triples propose rotations about an axis. Each candidate transformation $g$ is scored by attributed partial graph matching. Unmatched nodes are allowed and are labelled "possibly occluded" rather than counted as contradictions, so a scan missing one leg still votes for the mirror plane. The candidates that are accepted, together with their compositions, form a finite set $\hat G$. Because the skeleton is small, the search is cheap and its cost is independent of point count. This turns the unknown continuous-group problem into a finite one.
The third stage regularizes. We fit a neural signed distance field $f$ to the scan and minimize
$$ \mathcal{L} = \mathcal{L}_{\text{data}}(f) + \lambda_s \, \mathbb{E}_{g \sim \hat G}\, \mathbb{E}_{x}\, w_g \big(f(x) - f(g x)\big)^2 + \lambda_t \, \mathcal{L}_{\text{crit}}(f; S_{\hat G}). $$
The terms are defined as follows. $\mathcal{L}_{\text{data}}$ is the usual surface-fitting term. The expectation over $g$ uses a small random sample of $\hat G$, following the logarithmic-sample result. $x$ ranges over sampled 3D points. $w_g$ is the match confidence of $g$, so weak symmetries apply weak pressure. $\mathcal{L}_{\text{crit}}$ is a critical-component penalty in the style of the supervoxel loss. It compares the connected components and loops of the thresholded field against $S_{\hat G}$, the skeleton completed by mapping visible nodes onto their occluded partners. The data term dominates wherever the sensor actually observed the surface. The symmetry and topology terms act mainly in unobserved regions.
Beyond the method, we will measure break-even overhead directly. The ratio of detection time to reconstruction time, plotted against point count and noise level, will show when symmetry detection pays for itself. That analysis applies the planning and model-checking accounting to 3D perception.
Baselines
The naive reference is the same neural signed distance field fitted with only the data term: no topology term and no symmetry. It shows how much damage noise and occlusion do by default.
The strongest topology-only competitor is the Helmholtz–Hodge simply-connected reconstruction (Chen et al.). It reports CC=1 where competitors produce dozens of components, and roughly 15% lower Chamfer and Hausdorff error than its closest competitor on anatomical cross-section data. We run it on identical corrupted inputs with its published settings. Beating it means matching its component count while achieving lower Chamfer distance on occluded regions and correct Betti-1 counts on genus>0 shapes, where it is by construction wrong.
CrossSDF is the thin-structure competitor. It reports Chamfer distance 0.35 against 3.6 for OReX on Alveolis. We rerun it on the same slice data, add our terms to the same backbone, and give both the same iteration budget. We also test the same model with only $\lambda_t$ enabled, using a critical-component term in the style of Grim et al.. This model is the ablation that isolates symmetry's contribution.
We include two symmetry-detection baselines under the same wall-clock accounting. The first proposes reflection planes from principal axes of the raw points. The second runs exhaustive voting over point pairs on the raw cloud, the "detect on raw data" approach. For both we swap only the detector and keep the regularizer fixed. A further ablation replaces sampled averaging with exhaustive averaging over $\hat G$ to test the logarithmic-sample claim in this setting.
All methods share inputs, corruption seeds, the network backbone and a fixed hyperparameter search of equal size. Detection time is charged to total runtime. For point-cloud mesh completion on ScanObjectNN I found no reported result on Lacuna, so that comparison is established here rather than against a published figure.
Why not just…?
Why not just use a topology loss and skip symmetry? Topology losses can guarantee one connected component. They cannot say what the missing region looks like: an occluded leg reconstructed as a smooth stub still has correct Betti numbers. The topology-only ablation settles this. If it matches the full method on occluded-region Chamfer distance and on part-level recovery, the symmetry component is unnecessary and we will say so.
Why not just take the principal axes of the point cloud as the mirror plane? This is the real threat, because many man-made objects are scanned upright and nearly axis-aligned. Its known weakness is that occlusion shifts the center of mass and tilts the principal axes, and principal axes cannot represent rotational symmetries. We will report results stratified by occlusion fraction and symmetry type. If the principal-axes detector matches ours at realistic occlusion levels on reflection-symmetric objects, our contribution narrows to rotational and off-axis cases, and the paper will be framed that way.
Why not just detect symmetry exactly on the raw points or on a mesh graph? Exact automorphism search scales badly (Kunze et al.; Silva et al.). Under jitter, exact symmetry does not exist, so the detector returns the trivial group after spending its budget. Tolerance-based voting on raw points is the fair version of this alternative. The overhead-versus-noise curves will show where it breaks, and our prediction is that it breaks early.
Why not just train a completion network with heavy augmentation? Random jitter and dropout augmentations can themselves tear topology, as SinPoint documents. A learned model also offers no guarantee on shape categories it has not seen. We include a learned completion network as an extra reference on the synthetic benchmark. If it outperforms us in-distribution, we will test it on held-out categories, where per-scene symmetry should matter most.
Proposed Experiments
The primary benchmark is synthetic. We take complete watertight shapes from the CAD-style data used by BSP-Net and corrupt them with calibrated sensor models. The corruptions are Gaussian jitter, outliers, SinPoint-style smooth warps, and structured occlusion from simulated viewpoints that remove entire parts. The model is sensitive to noise, and we deliberately do not assume perfect sensing. Metrics are the number of connected components, Betti-0 and Betti-1 error, Chamfer distance computed separately on observed and occluded regions, false-symmetry rate, and detection time as a fraction of total time.
The second benchmark is the CrossSDF anatomical cross-section data. It tests thin branching structures, where symmetry is weak. There the confidence weights should shrink $\lambda_s$ automatically, and the method should do no worse than CrossSDF.
The third benchmark uses ScanObjectNN real scans. Ground-truth complete meshes are unavailable, so we report topological validity of the reconstructions. We also report classification accuracy when a PointNet++ is fed reconstructed rather than raw scans, alongside the 90.2% that SinPoint reports on this dataset.
Two checks decide the claim. First, on reflection- and rotation-symmetric shapes with 20–50% of the surface occluded, the full method must beat the topology-only ablation on occluded-region Chamfer distance, and the skeleton detector must beat raw-point voting on correct-symmetry recovery, with detection costing a small fraction of fitting time. Second, on deliberately asymmetric shapes, the false-symmetry rate must stay low and reconstruction quality must not degrade. Failing the first check means symmetry adds nothing over topology. Failing the second means the method is unsafe to deploy.
References
- Symmetry Reductions in Model Checking
- Exploiting Problem Symmetries in State-Based Planners
- Detecting and handling reflection symmetries in mixed-integer (nonlinear) programming and beyond
- Computing Game Symmetries and Equilibria That Respect Them
- Approximate Lifting Techniques for Belief Propagation
- Achieving Approximate Symmetry Is Exponentially Easier than Exact Symmetry
- Natural Graph Networks
- Guaranteed Simply Connected Mesh Reconstruction from an Unorganized Point Cloud
- CrossSDF: 3D Reconstruction of Thin Structures From Cross-Sections
- clDice - A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation
- Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation
- Towards Autonomous Grading in the Real World
- Fixed-Dimensional and Permutation Invariant State Representation of Autonomous Driving
- Practical Shuffle Coding
- Concentric network symmetry
- BSP-Net: Generating Compact Meshes via Binary Space Partitioning