3D data contain rich information about the full geometry of objects or scenes. Learning tasks on them have always been considered as hard ones in the computer vision community due to their extreme high dimensionality. Hence, latent representations of 3D geometries are often used to lower the data dimensionality for better parameterization and easier computation. In this report, we make a brief review on those latent representations obtained via different methods including classical ones and the emerging neural learning-based ones. Furthermore, the nowadays widely used deep learning methods have also been more closely investigated regarding their applications on various 3D data formats. The possibility of combing those two kinds of methods has also been addressed.