Mapper-GIN: Lightweight Structural Graph Abstraction for Corrupted 3D Point Cloud Classification

Robust 3D point cloud classification is often pursued by scaling up backbones or relying on specialized data augmentation. We instead ask whether structural abstraction alone can improve robustness, and study a simple topology-inspired decomposition based on the Mapper algorithm. We propose Mapper-GIN, a lightweight pipeline that partitions a point cloud into overlapping regions using Mapper (PCA lens, cubical cover, and followed by density-based clustering), constructs a region graph from their overlaps, and performs graph classification with a Graph Isomorphism Network. On the corruption benchmark ModelNet40-C, Mapper-GIN achieves competitive and stable accuracy under Noise and Transformation corruptions with only 0.5M parameters. In contrast to prior approaches that require heavier architectures or additional mechanisms to gain robustness, Mapper-GIN attains strong corruption robustness through simple region-level graph abstraction and GIN message passing. Overall, our results suggest that region-graph structure offers an efficient and interpretable source of robustness for 3D visual recognition.

💡 Research Summary

Abstract
The paper introduces Mapper‑GIN, a lightweight yet robust architecture for 3D point‑cloud classification under realistic corruptions. By converting a raw point set into an overlapping‑region graph using the Mapper algorithm (PCA lens, cubical cover, DBSCAN clustering) and then classifying this graph with a Graph Isomorphism Network (GIN), the method achieves competitive accuracy on the ModelNet40‑C benchmark while using only about 0.5 M parameters.

Motivation
Current state‑of‑the‑art point‑cloud models (PointNet++, DGCNN, PCT, etc.) rely on large backbones or extensive data augmentation to handle noise, density reduction, and geometric transformations. These solutions increase computational cost and often fail to capture the intrinsic structural invariants of objects. The authors ask whether a purely topological abstraction can provide robustness without heavy networks.

Background
Mapper, a tool from Topological Data Analysis (TDA), summarizes a high‑dimensional dataset by (1) projecting it through a low‑dimensional “lens,” (2) covering the lens space with overlapping sets, and (3) clustering points inside each pull‑back set. The resulting 1‑skeleton (the Mapper graph) encodes how local clusters overlap, revealing global connectivity such as branches and loops. Graph Isomorphism Networks (GIN) are the most expressive class of Graph Neural Networks (GNNs), matching the discriminative power of the Weisfeiler–Lehman test by using injective sum aggregation followed by an MLP.

Method

1. Topological abstraction – Each input cloud X⊂ℝ³ is projected by a PCA lens f_PCA(x)=Wᵀ(x−μ). The projected points are covered by a regular 3‑D grid (6 intervals per axis, 30 % overlap). For each grid cell U_i, the pull‑back set S_i = f_PCA⁻¹(U_i) is clustered with DBSCAN, yielding clusters C_{i,j}. Each cluster becomes a node n_{i,j} in the Mapper graph; edges connect nodes whose point sets intersect. This graph captures the coarse shape of the object while discarding fine‑grained geometric detail.
2. Node‑wise point encoder – For every node, the authors compute the centroid c_n and radius R_n, then normalize points to a local coordinate system. Two variants are explored: (a) Mapper‑GIN‑Base uses only the original coordinates as point descriptors, (b) Mapper‑GIN concatenates the original coordinates with the normalized local coordinates, forming a 6‑dimensional descriptor. A shared 1×1 convolutional MLP φ maps each descriptor to a point embedding, and max‑pooling over the points in the cluster yields the node feature z_n.
3. Graph classification with GIN – The node features are fed into a 4‑layer GIN. Each layer consists of a sum‑aggregation, a shallow MLP, GraphNorm, and ReLU. DropEdge and feature dropout are set to zero during training to keep the pipeline deterministic. After the final GIN layer, a global max‑pool aggregates node embeddings into a graph‑level vector g, which is passed through LayerNorm and a linear classifier to predict the object class.

Experiments
The authors evaluate on ModelNet40 and its corrupted variant ModelNet40‑C, which contains 15 corruption types (Gaussian noise, impulse noise, cutout, density reduction, rotations, shear, etc.) at five severity levels. Mapper‑GIN achieves the highest or near‑highest accuracy on the Noise and Transformation families, with mean accuracy improvements of 2–4 % over strong baselines such as DGCNN and PointTransformer, despite using an order of magnitude fewer parameters. Ablation studies show: (i) the number of intervals and overlap ratio affect the number of graph nodes but not robustness significantly; (ii) DBSCAN’s ε parameter is critical—too large merges distinct regions, too small creates many noisy nodes; (iii) adding the local normalized coordinates improves overall performance by roughly 1.5 %. The method is less robust to severe density‑reduction corruptions, where the Mapper graph itself changes dramatically, but still outperforms many heavy‑weight models.

Analysis and Insights

• Structural invariance: Rotations and non‑linear warps change point coordinates but preserve the connectivity pattern among overlapping clusters; consequently, the Mapper graph remains stable, and GIN can aggregate information across the unchanged topology.
• Noise mitigation: DBSCAN discards isolated outliers, and max‑pooling within each node reduces the influence of noisy points, leading to robustness against impulse and Gaussian noise.
• Parameter efficiency: By moving the heavy lifting from point‑wise feature extraction to a graph‑level abstraction, the model needs only a shallow MLP per point and a small GIN, dramatically cutting memory and compute requirements.
• Limitations: Objects with thin structures or many internal cavities may be over‑simplified by the fixed grid cover, causing loss of discriminative detail. The PCA lens is not fully rotation‑invariant, which could affect highly symmetric shapes. The current pipeline uses static hyper‑parameters for cover and clustering, limiting adaptability to varying scales of corruption.

Conclusion and Future Work
Mapper‑GIN demonstrates that a topologically motivated region graph, combined with an expressive yet lightweight GIN, can deliver strong corruption robustness with a fraction of the parameters of conventional point‑cloud networks. Future directions include exploring rotation‑invariant lenses (e.g., spherical harmonics), adaptive multi‑resolution covers, learnable clustering mechanisms, and integrating the Mapper construction into an end‑to‑end differentiable pipeline. Such extensions could further improve generalization to unseen corruption types and broaden applicability to real‑world 3D perception systems.