BIH: A novel satellite 3D point cloud generation method

The development of space technology and large-scale models have led to an increasing demand for 3D satellite data. However, current point cloud generation methods are not effectively applicable to satellite point cloud tasks due to the lack of 3D prior information. This deficiency results in the inability to generate reasonable geometric structures and the uniform point cloud distribution. Thus, we introduce a novel satellite point cloud generation method, called bidirectionally interactive homogeneity (BIH) model to expand satellite point clouds dataset. Through dual-interactive learning between generative adversarial network (GAN) and auto-encoder (AE), our method generates data support to solve the limitations faced by satellite large models. The interaction between GAN and AE of the BIH model markedly enhances the quality of satellites point clouds generation by leveraging and integrating more 3D prior information. Furthermore, our approach tackles the common issue of localized over-density in traditional generation algorithms through a homogeneity loss. We conduct extensive experiments on the NASA3D dataset, demonstrating that BIH model achieves superior performance compared to state-of-the-art methods. Specifically, it surpasses the best comparison method by 10.71% and 8.93% in the COV-CD and 1-NNA-CD, respectively. Moreover, we achieve double the data expansion and the new dataset named NASA3D-expand by training BIH model on NASA3D. Theoretically, BIH model has the capability to generate infinite satellite 3D point clouds, making it a substantial contribution to the development of large model techniques for satellite analysis.