Download
MeshCNN is a deep learning framework designed specifically for processing 3D triangular mesh data using convolutional neural networks. Unlike traditional CNNs that operate on images or voxel grids, MeshCNN performs convolution operations directly on the edges of mesh structures. This design allows the model to capture geometric relationships between mesh elements while preserving the underlying topology of 3D shapes. The framework introduces specialized layers such as edge-based convolution, mesh pooling, and mesh unpooling operations that enable hierarchical feature learning on mesh surfaces. These capabilities make the architecture well suited for tasks such as 3D object classification, segmentation, and geometric analysis. The project provides training pipelines, dataset preparation tools, and visualization utilities to support experiments with mesh-based neural networks.
Features
- Convolutional neural network architecture designed for triangular meshes
- Edge-based convolution operations tailored for geometric data
- Pooling and unpooling layers for hierarchical mesh feature learning
- Applications in 3D object classification and segmentation
- Training pipelines for experiments with mesh-based datasets
- Visualization tools for inspecting learned mesh transformations
Project Samples
