CS-Net: Contribution-Based Sampling Network for Point Cloud Simplification.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Point cloud sampling is vital for efficient processing in computer vision.
• Traditional methods like farthest point sampling lack task-specificity, limiting performance.
• Existing learning-based methods may not select the most relevant points or can produce duplicates.

Purpose of the Study:

• To introduce a novel contribution-based sampling network (CS-Net) for enhanced point cloud sampling.
• To address limitations of existing methods, including lack of task-specificity and duplicate point generation.
• To enable end-to-end training using differentiable approximations of Top-k operations.

Main Methods:

• CS-Net formulates sampling as a Top-k operation, using a differentiable approximation via optimal transport and entropy regularization.
• The network comprises feature embedding with spatial pooling, a cascade attention module, and a contribution scoring module.
• A novel spatial pooling layer and cascade attention mechanism refine feature representation and prioritize important points.

Main Results:

• CS-Net achieved state-of-the-art results on ModelNet40 and PU147 datasets for classification, registration, compression, and surface reconstruction.
• The method demonstrated high average precision for object detection on the KITTI LiDAR dataset.
• CS-Net effectively prioritizes relevant points, outperforming existing sampling techniques.

Conclusions:

• CS-Net offers a superior approach to point cloud sampling, enhancing performance across diverse downstream tasks.
• The proposed method overcomes key limitations of traditional and learning-based sampling techniques.
• CS-Net's effectiveness is validated across semantic and reconstruction-based tasks, as well as 3D object detection.