3D Generative Techniques and Visualization: A Brief Survey

The quality and quantity of data in the datasets is the key factor in producing accurate results for artificial intelligence applications. Real data is costly both from the time of gathering and from the labeling point of view. Moreover, the data property problem, the anonymization, and the representativeness are important factors that limit the dimension of the real datasets, making Synthetic Data Generation (SDG) the only alternative to produce large, high-quality datasets. The process of creating 3D synthetic data involves several steps, such as choosing the 3D model tool, creating the 3D model, applying texture and materials, setting up lighting, defining camera parameters, rendering the scene, augmenting data, adding depth and annotations, compiling the dataset, and performing validation and testing. Our paper explores the current landscape of 3D SDG, including generative methods, metrics, areas of application, existing packages to generate 3D data, and visualization of the generated data. The main objective is to focus on the specifics of 3D data, with an emphasis on the very recent state-of-the-art generative adversarial network techniques and assessment methods. We also discuss limitations of current 3D data generation techniques, challenges, and promising research directions.