LatEd: A Geometric Latent Vector Editor

Using a neural network approach, a shape may be compressed to a one-dimensional vector, the so-called latent dimension or latent vector. This latent shape dimension is examined in this paper. This latent vector of a shape is used to identify the corresponding shape in a database. Two types of networks are evaluated in terms of lookup accuracy and reconstruction quality using a database of Lego pieces. Even with small training set a reasonable robustness to rotation and translation of the shapes was achieved. While a human can interpret uncompressed data just fine, the compressed values of the network might be cryptic and thus offer no insight regarding the uncompressed input. Therefore, we introduce a latent dimension editor which allows the user to examine the geometry content of the latent vector and its influence on the decoded shape. The latent vector editor enables the visual exploration of the latent vector, by making changes to the latent vector visible in real-time via a 3D visualization of the reconstructed object.