AbstractIn visual culture, Generative Adversarial Networks (GANs) have been used to generate two-dimensional images, video, and three-dimensional forms. Whilst there are a relatively large number of conditional datasets of two-dimensional images, there are fewer datasets of three-dimensional objects available for training, evaluation, and practical use. In this paper, we introduce a synthetic dataset of 3D trees that provide novel geometric challenges for machine learning and describe our use of this dataset in training and qualitative evaluation via a public exhibition and survey. The 3D Tree Dataset was made using random variations of 76 bespoke tree templates based on art historical references to favor interesting, beautiful, and varied trunk and branch shapes. In generating and experimenting with this dataset, we wanted to know how a geometrically complex series of organic forms would challenge voxel-GANs compared to existing datasets comprising industrial objects such as chairs and cars. As an interdisciplinary project between visual arts and computer science, we used 3D printing and animation to visualize and communicate our outputs to non-specialist audiences. In this paper, we describe and share the 3D Tree Dataset, outline the process of generating the dataset, describe the GAN architecture and application used to test the dataset and discuss the results relative to our artistic and cultural goals.
Read full abstract