PointGL: A Simple Global-Local Framework for Efficient Point Cloud Analysis

TL;DR

PointGL introduces a simple yet effective global-local framework for point cloud analysis, replacing repeated per-point embeddings with a single Global Point Embedding per point and a lightweight Local Graph Pooling stage that builds small graphs to capture local geometry via edge features and max aggregation. The two-phase design dramatically reduces redundancy and computational cost while delivering strong accuracy, achieving state-of-the-art results on ScanObjectNN and competitive performance on ModelNet40, with substantial speedups and lower FLOPs/parameters. Extensive ablations validate the design choices, and robustness tests across ModelNet-C and ShapeNet-C demonstrate resilience to real-world corruptions. The approach also proves beneficial for segmentation and detection when integrated into existing backbones, highlighting practical impact for real-world 3D applications.

Abstract

Efficient analysis of point clouds holds paramount significance in real-world 3D applications. Currently, prevailing point-based models adhere to the PointNet++ methodology, which involves embedding and abstracting point features within a sequence of spatially overlapping local point sets, resulting in noticeable computational redundancy. Drawing inspiration from the streamlined paradigm of pixel embedding followed by regional pooling in Convolutional Neural Networks (CNNs), we introduce a novel, uncomplicated yet potent architecture known as PointGL, crafted to facilitate efficient point cloud analysis. PointGL employs a hierarchical process of feature acquisition through two recursive steps. First, the Global Point Embedding leverages straightforward residual Multilayer Perceptrons (MLPs) to effectuate feature embedding for each individual point. Second, the novel Local Graph Pooling technique characterizes point-to-point relationships and abstracts regional representations through succinct local graphs. The harmonious fusion of one-time point embedding and parameter-free graph pooling contributes to PointGL's defining attributes of minimized model complexity and heightened efficiency. Our PointGL attains state-of-the-art accuracy on the ScanObjectNN dataset while exhibiting a runtime that is more than 5 times faster and utilizing only approximately 4% of the FLOPs and 30% of the parameters compared to the recent PointMLP model. The code for PointGL is available at https://github.com/Roywangj/PointGL.

Figures (5)

• Figure 1: Comparing Accuracy and Speed Among Various Approaches on the ScanObjectNN Dataset. Our PointGL achieves the highest predictive accuracy and exhibits the fastest inference speed compared to alternative approaches.
• Figure 2: Comprehensive Architecture of PointGL. PointGL's architecture involves the extraction of hierarchical features from input point clouds through the stacking of multiple learning stages. Each stage initiates with a global point embedding phase, wherein feature embedding for individual points is conducted using residual MLP blocks. Following this, a local graph pooling phase captures and abstracts point-to-point relations into local representations by constructing a concise regional graph centered around each sampled point. The synergistic fusion of global point embedding and local graph pooling culminates in a coherent and efficient hierarchical framework for point cloud analysis.
• Figure 3: Salient geometric characteristics attained during the initial learning stage of PointGL on the ModelNet40 dataset. Each point's color corresponds to its received vote count, wherein a more intense red hue signifies a greater vote count, while a deeper blue shade indicates a lower count of votes.
• Figure 4: Visualization of part segmentation outcomes on the ShapeNetPart dataset. In contrast to PointNet++, PointGL's predictions exhibit a more robust alignment with the ground truth.
• Figure 5: Illustrative outcomes of 3D object detection on the KITTI dataset. Ground-truth and predicted objects are distinguished by red and green boxes, respectively. Integration of local graph pooling effectively recovers diminutive objects that were overlooked by the PV-RCNN baseline in intricate distant scenarios.