Abstract
Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling.
Introduction
In 3D computer vision, point clouds are often too dense for efficient processing, necessitating downsampling. While traditional methods like Farthest Point Sampling (FPS) are reliable, they are task-agnostic. Modern "learning-to-sample" approaches aim to identify the most important points for specific tasks, but they face two major hurdles: Edge Bias, where methods focus so heavily on sharp edges that they lose the global volume of the shape, and Static Strategies, where models apply the same sampling logic regardless of an object's unique geometry.
SAMBLE (Sparse Attention Map and Bin-based LEarning) addresses these issues by adaptively balancing the need for sharp local details with global shape uniformity. It ensures that sampled results capture the "bones" (edges) of an object without losing the "meat" (global uniformity) that defines its overall volume.
Methodology
SAMBLE introduces a shape-specific sampling pipeline consisting of three core components:
- Sparse Attention Map (SAM): To avoid the computational weight of global attention, SAMBLE utilizes a k-Nearest Neighbors (k-NN) approach to construct a sparse attention map. A key insight is that the frequency with which a point is chosen as a neighbor serves as a powerful geometric indicator, distinguishing between edge, "close-to-edge," and "non-edge" points.
- Point-wise Scoring: The model calculates a sampling score for every point. The authors found that a "column square-divided" indexing approach provided the best balance, preventing the model from over-focusing on edges.
- Bin-based Learning: Instead of a simple "Top-M" selection, SAMBLE partitions points into bins based on their scores. It uses learnable Bin Tokens to decide exactly how many points to draw from each bin for a specific shape. These bin boundaries are updated dynamically during training to adapt to the dataset's distribution.
Experimental Results
The authors evaluated SAMBLE on two industry-standard benchmarks: ModelNet40 for classification and ShapeNet-Part for part segmentation.
SAMBLE achieves state-of-the-art (SOTA) performance across multiple metrics:
- Classification Accuracy: On ModelNet40, SAMBLE reached 94.2% Overall Accuracy, surpassing previous methods like APES (93.8%).
- Robustness in Few-Point Sampling: The method is remarkably robust at extreme compression. When sampling down to just 32 points, SAMBLE maintained 89.45% accuracy, while competing methods saw much steeper performance drops.
- Segmentation Performance: It consistently outperformed PointNeXt and APES on intermediate downsampled sub-point clouds, proving its utility for downstream geometric understanding.
Qualitatively, visualizations demonstrate that while other models might reduce a chair to just its corners, SAMBLE retains points on the seat and back, maintaining the global uniformity necessary for tasks like upsampling and human recognition.
Conclusion and Future Work
SAMBLE demonstrates that geometric importance is not just about local curvature, but about how points relate to each other within an attention mechanism. By using representative bin-based sampling, the model ensures it picks a mix of edge and surface points that preserve an object's identity.
While SAMBLE provides a superior foundation for sampling, future work could involve designing more meticulous upsampling architectures specifically tailored for this method. Additionally, extending this shape-specific strategy to complex scene-level environments remains a promising and challenging next step.