Hybrid Tensor-Gaussians Representation with Space-Aware Pruning

Recently, significant strides have been made in the realm of 3D Gaussian Splatting, marking notable advancements in novel view synthesis. However, due to randomness in the optimization process, Gaussians often freely disperse throughout the three-dimensional space, yielding substantial redundancies. By carefully regulating the opacity and scale, various pruning methodologies have effectively curtailed the proliferation of Gaussian distributions, yet they concurrently excised numerous pivotal Gaussians. As a remedy, we introduce a new solution, namely TSGS, leveraging structured information within tensor grids to constrain the unstructured growth of Gaussian distributions. During the optimization, we further integrate a space-aware mask, intelligently situating Gaussian distributions within essential spatial locales and preserving many invaluable distributions predicated on spatial cues. In addition, a joint optimization of opacity and color is established, which effectively controls the optimization direction of the model and reduces randomness. Extensive experiments demonstrate that our method outperforms state-of-the-art methods, effectively reducing redundant Gaussians while delivering higher-quality rendering, with much lower memory consumption.