Abstract
In this paper, we diagnose deep neural networks for 3D point cloud processing to explore utilities of different network architectures. We propose a number of hypotheses on the effects of specific network architectures on the representation capacity of DNNs. In order to prove the hypotheses, we design five metrics to diagnose various types of DNNs from the following perspectives: information discarding, information concentration, rotation robustness, adversarial robustness, and neighborhood inconsistency. We conduct comparative studies based on such metrics to verify the hypotheses. We further use the verified hypotheses to revise architectures of existing DNNs and improve their utilities. Experiments demonstrate the effectiveness of our method.
Introduction
Deep Neural Networks (DNNs) for 3D point clouds have achieved significant performance gains, yet their architectural design remains largely heuristic. While researchers know that certain modules improve performance, the underlying why—the specific utility a module provides—often remains opaque. This work aims to bridge the gap between intermediate-layer architectures and their specific utilities, such as adversarial robustness or neighborhood consistency. By establishing a framework of Verifiability (proving a module's function) and Predictability (using that proof to enhance other models), the authors move point cloud processing toward scientific rigour.
Diagnostic Framework
To quantify the internal logic of point cloud networks, the authors developed five diagnostic metrics grounded in information theory and entropy-based quantification:
- Information Discarding: Measures input information loss during forward propagation.
- Information Concentration: Evaluates the model's focus on task-related "foreground" points versus "background" noise.
- Rotation Robustness: Assesses if internal logic remains consistent regardless of object orientation.
- Adversarial Robustness: Measures resistance to small, malicious input perturbations.
- Neighborhood Inconsistency: Checks for similar importance assignments to adjacent points with similar geometric contexts.
Architectural Hypotheses
The study focuses on four specific intermediate-layer architectures to test hypotheses regarding their utilities:
- Architecture 1: Uses local density information to reweight features, hypothesized to improve adversarial robustness.
- Architecture 2: Uses local 3D coordinates to reweight features, hypothesized to drive rotation robustness.
- Architecture 3: Concatenates multi-scale features, hypothesized to improve both adversarial robustness and neighborhood consistency.
- Architecture 4: Computes orientation-aware features to enhance rotation robustness.
Experimental Results and Predictability
The researchers evaluated seven popular DNNs (e.g., PointNet, PointNet++, DGCNN) across benchmarks including ModelNet40, ShapeNet, and 3D MNIST. To test information concentration, a custom version of ModelNet40 with added "background" points was also utilized.
A key contribution is the "predictable" improvement of existing networks. By integrating verified modules into baseline models, the authors observed consistent improvements ($\Delta$) across metrics and datasets. For example, adding Architecture 2 (local coordinate reweighting) to PointNet++ improved its rotation robustness score by 3.248 on ModelNet40. Similarly, Architecture 3 (multi-scale) improved RSCNN's neighborhood consistency by 0.088 and adversarial robustness by 0.036.
Conclusion
This research provides a systematic way to interpret and improve 3D point cloud networks. By verifying the specific utilities of architectural modules, designers can move beyond trial-and-error toward a more predictable engineering process. Future work may extend these diagnostic metrics to more complex architectures like transformers and broader tasks such as 3D segmentation.
Summary updated