Advanced Feature Learning on Point Clouds Using Multi-Resolution Features and Learnable Pooling

Abstract

Existing point cloud feature learning networks often learn high-semantic point features representing the global context by incorporating sampling, neighborhood grouping, neighborhood-wise feature learning, and feature aggregation. However, this process may result in a substantial loss of granular information due to the sampling operation and the widely-used max pooling feature aggregation, which neglects information from non-maximum point features. Consequently, the resulting high-semantic point features could be insufficient to represent the local context, hindering the network's ability to distinguish fine shapes. To address this problem, we propose PointStack, a novel point cloud feature learning network that utilizes multi-resolution feature learning and learnable pooling (LP). PointStack aggregates point features of various resolutions across multiple layers to capture both high-semantic and high-resolution information. The LP function calculates the weighted sum of multi-resolution point features through an attention mechanism with learnable queries, enabling the extraction of all available information. As a result, PointStack can effectively represent both global and local contexts, allowing the network to comprehend both the global structure and local shape details. PointStack outperforms various existing feature learning networks for shape classification and part segmentation on the ScanObjectNN and ShapeNetPart datasets, achieving 87.2% overall accuracy and instance mIoU.