Conditional GAN for Point Cloud Generation

Zhulun Yang, Yijun Chen, Xianwei Zheng, Yadong Chang, Xutao Li; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 3189-3205

Recently, 3D data generation problems have attracted more and more research attention and have been addressed through various approaches. However, most of them fail to generate objects with given desired categories and tend to produce hybrids of multiple types. Thus, this paper proposes a generative model for synthesizing high-quality point clouds with conditional information, which is called Point Cloud conditional Generative Adversarial Network (PC-cGAN). The generative model of the proposed PC-cGAN consists of two main components: a pre-generator to generate rough point clouds and a conditional modifier to refine the last outputs with specific categories. To improve the performance for multi-class conditional generation for point clouds, an improved tree-structured graph convolution network, called BranchGCN, is adopted to aggregate information from both ancestor and neighbor features. Experimental results demonstrate that the proposed PC-cGAN outperforms state-of-the-art GANs in terms of conventional distance metrics and novel latent metric, Frechet Point Distance, and avoids the intra-category hybridization problem and the unbalanced issue in generated sample distribution effectively. The results also show that PC-cGAN enables us to gain explicit control over the object category while maintaining good generation quality and diversity. The implementation of PC-cGAN is available at https://github.com/zlyang3/PC-cGAN.