Where Does Topology Actually Survive Corruption? A Per-Corruption Stability Audit for Point Cloud Recognition

Topological data analysis summarizes a shape by its holes and connected pieces, and the scale at which each one appears and disappears. Its advocates often call these summaries "robust to noise." The evidence behind that claim is thin. It comes from toy loops with sprinkled outliers, MNIST with dropped pixels, and point clouds downsampled to 128 points. Separately, the robustness community has built standardized corruption suites. ModelNet40-C, for example, applies 15 controlled distortions (occlusion, LiDAR scan patterns, background noise, shear, smooth warps, and others) at five severities to 3D shapes. Under these suites, strong point cloud classifiers roughly triple their error. Nobody has used these suites to check where topological summaries actually stay stable, and the theory gives no clear answer either. A smooth warp leaves the true holes of an object intact but can move every pairwise distance. A single background point can plug a hole outright.

This note proposes to measure, before training anything, how far each corruption moves an object's topological summary relative to how far apart different classes' summaries are. We then test whether that cheap ratio predicts on which corruptions a topological feature branch helps a deep classifier, and on which it is dead weight. If the prediction holds, practitioners get a principled way to decide when topology is worth its cubic compute cost. If it fails, the field learns that the "noise robustness" of robust filtrations does not carry over from toy settings to realistic corruption.

Core Claim

A persistence diagram is the set of (birth scale, death scale) pairs for every hole and component found as balls grow around the points. A filtration is the rule that decides when each point and edge enters that growth process. For a given filtration $F$, corruption type $c$ and severity $s$, define the sensitivity ratio

$$S_F(c,s) = \frac{\operatorname{median}_x\, W_2\big(D_F(x),\, D_F(c_s(x))\big)}{\operatorname{median}_{x,x' \text{ in different classes}}\, W_2\big(D_F(x),\, D_F(x')\big)},$$

where $D_F(x)$ is the diagram of clean shape $x$ under filtration $F$, $c_s(x)$ is its corrupted version, and $W_2$ is the 2-Wasserstein distance between diagrams. When $S_F$ is well below 1, the corruption moves an object's topology less than classes differ, so a classifier that sees the diagram should keep its decision. When $S_F$ approaches or exceeds 1, the topological branch carries corrupted evidence and may hurt.

Our claim is that ranking the 15 corruptions by $S_F$ predicts, with high rank correlation, the per-corruption accuracy gain from fusing a topological branch with a standard backbone. We also claim that the ranking differs systematically by filtration. Plain distance-based filtrations should fail on background and impulse noise. Density-weighted filtrations should repair noise but not occlusion. Smooth deformations, which preserve true homology but distort distances, should sit in between, depending on displacement magnitude.

Relation to Prior Work

Three topological methods motivate the audit, and each supplies a filtration we reuse. PLLay builds a differentiable layer over distance-to-measure diagrams. Distance-to-measure replaces "distance to the nearest point" with an average distance to the $k$ nearest points, so isolated outliers enter late. PLLay reports 95.0% on ORBIT5K versus 70.8% for PointNet, but it tested robustness only on MNIST with random pixel drops and dots. Filtration learning for point clouds trains a network to assign per-point filtration weights. On 128-point noisy ModelNet10, it lifts PointNet from 64.3% to 68.8% under a single ad-hoc noise protocol. The Λ-filter handles noise and uneven density together and beats distance-to-measure on Bone Scripts (89% vs 80%). Its authors state that it struggles with twisting and shearing. All three report noise robustness on one or two hand-built perturbations. We take their filtrations unchanged and ask a question none of them asks: across a calibrated taxonomy and five severities, where does each one hold?

Multiparameter persistence images filter by scale and density at once. They are a natural fourth filtration, but their slice-matching is unstable near feature merges, so we include them only on a subset where runtime allows. Topological analysis of scalar fields with outliers supplies the stability theory that makes distance-to-measure's behaviour predictable. Our ratio is an empirical analogue of those bounds, normalized by class separation.

On the benchmarking side, ModelNet40-C supplies the instrument. It also supplies a constraint: corruption robustness differs sharply by architecture. Ball-query neighborhoods resist background noise better than k-nearest-neighbor ones, and PointNet resists density changes but not rotation. Any topological gain must therefore be measured per backbone. The original common-corruptions benchmark warns against training on the test corruptions, and we follow that rule. SMART-PC is the closest bridge in the file. It argues that internal skeletal structure (medial spheres) is more stable than surface points under ModelNet40-C, ScanObjectNN-C and ShapeNet-C, and adapts at test time. It never computes persistent homology or checks whether topological invariants survive. We differ by measuring topological stability directly and by predicting where it pays off, rather than adding another heuristic structural branch.

Research Program

The first stage is the audit. We draw a stratified sample of about 40 shapes per class from ModelNet40 and pair each with all 75 corrupted versions in ModelNet40-C. We reduce each cloud to 256 landmarks by farthest-point sampling, which keeps the cubic cost of persistence tractable. We then compute connected-component and loop diagrams under four filtrations: Vietoris–Rips (plain distances), distance-to-measure at three values of $k$, the Λ-filter, and the learned filtration of Nishikawa et al. trained on clean data only. This yields a 15×5 sensitivity map per filtration. We repeat the audit on ScanObjectNN-C, where real scans carry background clutter, to check that the map is not a quirk of synthetic CAD meshes.

The second stage is the intervention. We add a topological branch to three backbones of different neighborhood types. PointNet has no neighborhoods. DGCNN uses k-nearest-neighbor neighborhoods. PointMLP, built from its official code, uses residual local aggregation. The branch vectorizes each diagram as a persistence landscape, following PLLay, and concatenates it to the global feature. All models train on clean ModelNet40 with the same standard augmentations. No ModelNet40-C corruption is ever seen in training, and no test-time adaptation is used unless stated.

The third stage tests the prediction. We correlate the per-corruption accuracy gain with $S_F(c,s)$ computed in stage one. We also test a deliberately pre-registered hypothesis about smooth deformations, free-form deformation and radial-basis warps. Because these are homeomorphisms, true holes persist, so long-lived features should survive at low severity. They should collapse only when displacement approaches the gap between a loop's birth and death scales. If this holds, ModelNet40-C's transformation family is not uniformly "hostile to topology," as the gap analysis in the file assumes.

Baselines

The naive reference is each backbone trained on clean data and evaluated on ModelNet40-C with no topological branch. ModelNet40-C's authors report a 26.1% average error for their baseline setting, and that defines the floor any added component must beat.

The strongest existing method in the file is PointCutMix-R augmentation combined with test-time entropy minimization (TENT), which ModelNet40-C reports reduces average error from 26.1% to 15.4%. We rerun it on our backbones with the same training epochs and the same hyperparameter search budget as our topological variants, about 20 configurations each, selected on a clean validation split. The topological branch is meant to complement, not replace, this method. "Beating" it therefore means the combination of PointCutMix-R, TENT and the topological branch lowers average error below the pair's own result, with the gain concentrated on the corruptions our map predicts.

On ScanObjectNN-C we report SMART-PC's 47.4% mean accuracy as the reference for structure-aware test-time methods. Matching or exceeding it with a branch that does not adapt at test time would be a meaningful result. Falling short while still showing per-corruption gains where predicted would still support the core claim.

Two ablations isolate the key ingredient. The first, "density weighting without topology," removes points whose distance-to-measure value exceeds a threshold before feeding the backbone, with the same $k$ and no persistence computation. The second, "topology without robustness," uses a plain Rips branch. If the first matches the full branch, the benefit comes from density filtering, not from homology.

Why not just…?

Why not just remove outliers before the network? This is the most serious threat, because distance-to-measure is essentially a k-nearest-neighbor density score, and statistical outlier removal does the same thing cheaply. The density-weighting ablation tests this directly. We expect preprocessing to match the topological branch on background and impulse noise. Preprocessing cannot recover a loop broken by occlusion or cutout, and it cannot tell a smoothly warped shape from a differently shaped one. If the ablation matches the full branch on every family, we will report that topology adds nothing beyond density filtering on these benchmarks. That is a useful negative result, and the sensitivity map remains valid as a diagnostic.

Why not just use augmentation plus test-time adaptation, which already cuts error to 15.4%? That result shows strong average gains, but averages hide per-corruption behaviour, and the reported rotation benefit of TENT (27.1% over augmentation alone) shows the gains are uneven across families. Our experiments stack the topological branch on top of this method. The question is whether residual errors cluster on low-$S_F$ corruptions that topology can address.

Why not just switch to ball-query neighborhoods, which already resist background noise? ModelNet40-C shows ball query beats k-nearest-neighbor aggregation on background noise, so part of any topological gain on PointNet or DGCNN may simply be a neighborhood-choice effect. Including PointMLP and reporting gains per backbone settles this. If the gain on background noise vanishes on ball-query-style backbones while persisting on occlusion, the effect is not reducible to neighborhood choice.

Proposed Experiments and Decisive Checks

The primary metric is per-corruption, per-severity error. From it we report mean corruption error normalized to clean-trained PointNet, relative corruption error against clean accuracy, and decay curves of accuracy versus severity for each family. Results are averaged over three seeds, given the run-to-run instability that the PointMLP repository itself notes. The audit reports $S_F$ with bootstrap intervals over shapes.

The decisive check is the rank correlation, across the 75 corruption–severity cells, between $S_F$ from the training-free audit and the accuracy gain of the matched topological branch. A correlation above roughly 0.6 that holds on two of three backbones would show that topological sensitivity is predictable before training. A correlation near zero would refute the core claim.

The second check is the homeomorphism hypothesis. Long-lived loop features should be retained (bottleneck shift below half the median loop lifetime) for free-form and radial-basis warps at severities 1–3, and lost for occlusion at the same severities.

The third check is cost. We report wall-clock time for the topological branch at 256 landmarks against the accuracy it buys per corruption family. This tells a practitioner when the audit says the cubic cost is worth paying.