Quantized 2D Gaussian Splatting for Memory Efficient 3D Reconstruction and Novel View Synthesis

2D Gaussian Splatting is an approach to model and reconstruct geometrically accurate radiance fields from multi-view images. The method is in turn derived from 3D Gaussian Splatting, an approach that revolutionized radiance field reconstruction but which itself has critical issues by failing to always accurately represent surfaces due to the multi-view and inconsistent nature of 3D Gaussians. The main difference in the new approach lies precisely in the use of 2D Gaussians as a primitive for 3D volume representation instead of the 3D Gaussian, which implicitly provides view-consistent geometry and is capable of modeling intrinsic surfaces. The system is accompanied by a perspective-accurate 2D splatting process using ray-splat intersection and rasterization, with the aim of accurately recovering thin surfaces and achieving stable optimization. 2DGS has competitive appearance quality, high training speed, and real-time rendering. The purpose of this thesis is to go for reducing the memory storage of the 3D reconstruction obtained by 2DGS, keeping training and rendering times aligned with the original version, through three main elements: a vector quantization of Gaussian parameters based on -means; a regularization function that acts only on opacity to go for reducing the number of Gaussians with too low an opacity value; a modification to the CUDA kernel code executed in the rasterization phase for the removal of points and consequently Gaussians. Similar approaches have already been developed with regard to 3DGS, and as can be seen from the results obtained using the DTU dataset benchmark, this approach also appears to be valid in the context of 2DGS