Decoding 3D Geometry: Deep Networks for Mesh Classification

Three-dimensional (3D) meshes are a fundamental representation for objects in applications spanning computer graphics, gaming, and simulation. While convolutional neural networks (CNNs) have become the de facto standard for processing 2D images, a standard deep learning technique for 3D meshes remains an open challenge due to their irregular structure. Developing robust, mesh-native networks is critical for tasks like automatic content creation and style transfer, as it eliminates the overheads and potential information loss associated with converting meshes to other formats like point clouds or volumetric grids.This paper provides an empirical analysis of deep learning models for 3D mesh classification. We benchmark and fine-tune two prominent architectures, MeshCNN and Dynamic Graph CNN (DGCNN), on standard datasets, systematically investigating the impact of their core hyperparameters. Building on this analysis, we propose and evaluate a novel hybrid model that combines the complementary feature extraction capabilities of MeshCNN and DGCNN. We analyze its performance against the individual models to understand the benefits and challenges of feature fusion in this domain.