We introduce UNIST, the first deep neural implicit model for general-purpose, unpaired shape-to-shape translation, in both 2D and 3D domains. Our model is built on autoencoding implicit fields, rather than point clouds which represents the state of the art. Furthermore, our translation network is trained to perform the task over a latent grid representation which combines the merits of both latent-space processing and position awareness, to not only enable drastic shape transforms but also well preserve spatial features and fine local details for natural shape translations. With the same network architecture and only dictated by the input domain pairs, our model can learn both style-preserving content alteration and content-preserving style transfer. We demonstrate the generality and quality of the translation results, and compare them to well-known baselines. Code is available at https://qiminchen.github.io/unist/.

In the application of 3D object retrieval we search for 3D objects similar to a given query object. When a user searches for a certain class of objects like 'planes' the results can be unsatisfying: Many object variations are possible for a single class and not all of them are covered with one or a few example objects. We propose a meta model representation which corresponds to a procedural model with meta-parameters. Changing the meta-parameters leads to different variations of a 3D object. For the meta model generation a single object is constructed with a modeling tool. We automatically extract a procedural representation of the object. By inserting meta-parameters we generate our meta model. The meta model defines a whole object class. The user can choose a meta model and search for all objects similar to any instance of the meta model to retrieve all objects of a certain class from a 3D object database. We show that the retrieval precision is significantly improved using the meta model as retrieval query.