Abstract Standard point cloud architectures rely on max-pooling to achieve permutation invariance, a process that discards significant geometric detail. We introduce SoftPoolNet, a novel architecture that replaces max-pooling with a "soft pooling" layer. By sorting feature activations, SoftPoolNet retains a richer descriptor while remaining permutation-invariant. Combined with regional convolutions and entropy-based losses, this approach achieves state-of-the-art performance in 3D point cloud completion and classification, particularly in reconstructing fine-grained details from partial scans.
Introduction
Point clouds, the raw output of LiDAR and depth sensors, consist of unorganized $(x, y, z)$ coordinates. Processing them requires architectures that are invariant to the input order. Pioneering models like PointNet utilize max-pooling to extract global features. However, max-pooling is inherently lossy, retaining only the maximum activation per channel. This "bottleneck" makes it difficult for decoders to accurately reconstruct missing parts of a shape, as local geometric nuances are lost during encoding.
Methods
SoftPoolNet addresses the information loss of max-pooling through three primary innovations:
- Soft Pooling: Instead of selecting a single maximum value, features are sorted based on their activations. The top $N_r$ highest-scoring features for each channel are retained to form a structured "softpool feature" matrix ($F^*$). This deterministic sorting ensures permutation invariance while preserving a significantly larger "memory" of the input shape.
- Regional Convolutions: The decoder uses a specialized operator that slides kernels within the organized regions of the softpool feature. This allows for effective upsampling and interpolation of local features, enabling the generation of dense point clouds (e.g., 16,384 points) from coarse representations.
- Entropy-Based Loss Functions: To guide the learning of meaningful regions, the authors employ Regional Entropy losses. These encourage even point distribution across regions (inter-regional) and high confidence for point-to-region assignments (intra-regional). A Boundary Loss is further used to prevent regional overlap and reduce surface noise.
Results and Discussion
The model was evaluated on ShapeNet (completion), KITTI (real-world LiDAR), and ModelNet (classification). SoftPoolNet demonstrates superior performance in reconstructing complex and rare geometries, such as six-legged tables, which standard models often over-simplify.
On the ShapeNet benchmark, SoftPoolNet achieved a Chamfer Distance (CD) of 5.94, a substantial improvement over the previous state-of-the-art PCN (9.64). In classification tasks, it reached 92.28% accuracy on ModelNet40. Furthermore, the model is highly efficient, with an inference time of approximately 0.04 seconds per object, making it suitable for real-time applications.
Conclusion
SoftPoolNet demonstrates that retaining more information through sorting-based pooling significantly enhances the capability of point cloud networks. By "ordering the unordered," the architecture enables standard convolutional operations to be applied effectively to 3D data. Future work will explore the application of softpool descriptors to scene segmentation and automated hyperparameter selection for region counts.
Summary updated