Abstract
SoftPoolNet introduces a novel "soft pooling" layer and "regional convolutions" to overcome the information bottleneck of standard max-pooling in point cloud processing. By retaining multiple high-activation features and organizing them into learnable regions, the architecture achieves state-of-the-art performance in 3D object completion and classification, producing significantly finer geometric details than previous methods.
Introduction
Point clouds are the fundamental data format for 3D sensors like LiDAR, representing objects as unordered sets of coordinates $(x, y, z)$. While flexible, their unorganized nature makes them difficult for standard Deep Learning architectures to process. The seminal PointNet architecture addressed this by using a max-pooling layer to achieve permutation invariance. However, max-pooling is inherently lossy; it discards all but the single highest activation for each feature dimension. This creates a significant information bottleneck that limits a network's ability to reconstruct fine details or recognize complex, rare geometries, such as a table with six legs instead of the standard four.
Methods
To break this bottleneck, SoftPoolNet introduces two primary architectural innovations: Soft Pooling and Regional Convolutions.
Soft Pooling
Instead of keeping only the maximum activation, SoftPoolNet sorts feature vectors based on their activation values. It selects the top $N_r$ features for each of the $N_f$ feature dimensions. This preserves a much richer set of "high-scoring" descriptors while maintaining permutation invariance through the sorting process. This allows the network to capture a broader distribution of the input geometry rather than a single peak activation.
Regional Convolutions
The features preserved by soft pooling are organized into "regions." The authors introduce a regional convolution operator that slides a kernel within these regions. This allows the decoder to act like a traditional CNN, refining local details and upsampling the point cloud from a coarse representation (e.g., 256 points) to a high-resolution output (e.g., 16,384 points).
Loss Functions
The training process is guided by a suite of specialized loss functions:
- Regional Entropy: Encourages the network to distribute points evenly across regions (inter-regional) while ensuring each point belongs strongly to a specific region (intra-regional).
- Boundary Loss: Minimizes the overlap between adjacent regions to prevent "noisy" reconstructions where different patches are stacked on top of each other.
- Preservation Loss: Uses Earth-Mover’s Distance (EMD) to ensure the features extracted by the encoder accurately represent the original input distribution.
Key Insights and Ideas
One of the most striking emergent properties of SoftPoolNet is unsupervised segmentation. The network learns to segment objects into meaningful parts (e.g., the wings of a plane or the legs of a chair) entirely through the regional entropy loss, without any human-provided part labels.
Furthermore, because SoftPoolNet retains more information, it avoids the "regression to the mean" problem common in point cloud completion. While other models might complete a partial six-legged table as a standard four-legged one, SoftPoolNet has the capacity to represent and reconstruct rare or non-standard geometries.
Experimental Results
The model was evaluated on standard benchmarks including ShapeNet, KITTI, ModelNet10/40, and PartNet.
State-of-the-Art Completion
On the ShapeNet dataset, SoftPoolNet achieved a Chamfer Distance of 5.94, significantly outperforming established baselines like PCN (9.64) and FoldingNet (10.07). Qualitatively, the method produces much sharper edges and more complete surfaces than competing voxel-based or point-based methods.
Classification and Efficiency
SoftPoolNet reached 92.28% accuracy on ModelNet40, surpassing models like PointNet++ and RS-DGCNN. Despite the increased complexity of the features, the model is highly efficient, boasting an inference time of 0.04 seconds per sample—nearly 3x faster than PCN and 8x faster than AtlasNet.
Conclusion and Future Work
SoftPoolNet demonstrates that sorting-based pooling is a viable and more informative alternative to max-pooling for handling unordered sets. While it excels at reconstructing thin structures and rare geometries, extremely fine details remain a challenge compared to mesh-based approaches. Future work may explore dynamic region assignment or applying these regional convolutions to large-scale scene completion.
Summary updated