DeepCloud (2018)
DeepCloud depicts a symbiotic design future where designers can curate their own design culture on the cloud, while an artificial intelligence system defines the appropriate design space and operations accordingly. In its current configuration, DeepCloud is a data-driven modeling system that enables the user to quickly generate innovative and unexpected objects of any class in its database – such as cars, chairs, tables, and hats. It learns the common characteristics of these classes and enables the user to manipulate them in a meaningful way.
All objects are represented as sets of points in space. DeepCloud uses an autoencoder to learn how to compress them into a low-dimensional latent feature space and how to rebuild them back to the original space of the point clouds. In this process, it becomes a generative design system in which new objects (represented by point clouds) can be synthesized through the manipulation or combination of the vectors in the latent feature space.
DeepCloud enhances the experience of synthesizing new objects in real-time with a web- interface integrated with a MIDI mixer as the input device. There are two operations available for design exploration: 1) interpolation, and 2) feature manipulation. In the former, the user selects a set of objects in the menu and can interact with the mixer to combine all its features, generating new hybrid objects. In the latter, the interaction with the mixer transforms each dimension of the feature space, starting with a known or random point cloud. As each dimension is associated with certain characteristics of the class, its manipulation results in non-linear geometric transformations that are potentially meaningful, such as growing an armrest, adding an opening to the back of a chair, adding a spoiler to the trunk of a car, etc.
Disclaimer: DeepCloud is developed based on the Latent 3D Point project by Achlioptas et. al.
- Achlioptas, Panos, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. “Learning Representations and Generative Models for 3D Point Clouds.” (2018).
- https://github.com/optas/latent_3d_points
Acknowledgment: The authors are grateful to Prof. Eunsu Kang and Prof. Barnabas Poczos for supervising this project in their Art and Machine Learning course at Carnegie Mellon University, School of Computer Science. We would like to thank Shenghui Jia for his great contribution to the 3D- printing process. We would like to express our gratitude to CMU’s Computational Design Lab (CodeLab) for its generous