For Stanford’s Convolutional Neural Networks class, I worked on a project to perform shape classification using a variety of architectures. Using 3D shapes from the ShapeNet project, they are first centered, rescaled and resampled to occupancy voxel grids of dimension 30 x 30 x 30. We implement 3D convolutional neural networks operating directly on these three dimensional volumes, as well as 2D convolutional networks operating on a learned embedded from the 3D model. Results from both methods are state-of-the-art compared to existing, published methods.

CS231N Paper

CS231N Poster