LBWGAN: Label Based Shape Synthesis from Text with WGANs

In this work, we purpose a novel method of voxel-based shape synthesis, which can build a connection between the natural language text and the color shapes. The state-of-The-Art method use Generative Adversarial Networks (GANs) to achieve this task and some achievements have been made with it. It is a very advanced framework on this subject but the state-of-The-Art method significantly ignores the role of the class labels. Labels can guide shape synthesis because shapes in different labels have different characteristics. Therefore, this work attempts to create a deeper connection between the labels and the generated results. It based on a new structure and lets the labels guide the shape synthesis work. A key idea is to establish a new set of relationships outside the generator and discriminator to guide the training process. This paper introduces an independent class classifier in the new structure and makes it grow together with the generator to make the generated results have more distinctive class features. Experiments show that our method has a more exquisite performance on the synthesis of complex shapes, performing more realistic, and has better performance in structural integrity. Besides, our approach can extract the implied shape messages from the descriptions to realize shape synthesis.