DGONN: Depthwise Dynamic Graph Overparameterized Neural Network for 3D Point Cloud Object Recognition

Rapid advancement in 3D point cloud object recognition is crucial for robotics, autonomous driving, and augmented reality applications. The traditional methods, including PointNet and its successors, though effective in handling unordered point cloud data, need help capturing local structures accurately and efficiently. This paper introduces a novel architecture, the Depthwise Dynamic Graph Overparameterized Neural Network (DGONN), which enhances point cloud object recognition by integrating graph-based features with overparameterized networks. Our method leverages local geometric formations through a neighborhood graph. It performs operations similar to convolutions, utilizing edge convolution (EdgeConv) and depthwise overparameterized convolution (DO-Conv) for dynamic graph updates and efficient feature representation. The proposed DGONN architecture dynamically updates the graph structure with each layer, allowing for adaptive learning and improved performance in 3D object recognition tasks. Through extensive experiments, DGONN demonstrated superior performance over state-of-theart methods across various metrics on the ModelNet40 and ScanObjectNN datasets with accuracy scores of 92.9% and 78.3%, respectively. This performance highlights its effectiveness in preserving dense spatial relationships and patterns within point cloud data. Future work focuses on making the system faster and more efficient by improving the model's ability to work well with different types of point cloud data, even in challenging conditions like outdoor scenes, and incorporating new features like texture.