GCN-Based Objects Understanding with 2D to 3D Point Cloud Reconstruction
Hamid Ashfaq, Ahmad Jalal
Abstract
Estimating and generating 3D models from single images is a challenging problem that has gained considerable attention in computer vision and artificial intelligence. This article focuses on 3D reconstruction using standard and state-of-the-art methods, with previous research exploring single-angle and multi-view approaches. Various representations of 3D data, such as meshes, voxels, and point clouds, are discussed. The proposed semi-neural network system combines neural networks, image processing filters, and machine learning algorithms to extract features from single images. Three key features, including semantic segmentation, image depth, and surface normal, aid in estimating the 3D structure. The resulting 3D models are represented as point clouds, facilitating model quality analysis. This article evaluates and assesses the system using three benchmark datasets: ShapeNet, ModelNet10, and ObjectNet3D. ShapeNet achieved an accuracy of 95.41% with a chamfer distance of 0.00098, ModelNetl0 achieved an accuracy of 94.74% with a chamfer distance of 0.00132, and ObjectNet3D achieved an accuracy of 95.53 % with a chamfer distance of 0.00091. The proposed system's visualization results for various classes surpass those of standard methods, demonstrating outstanding performance.