BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis

• DOI: 10.48550/arxiv.2411.08508
• Publication Date: 2024-11-13
• Authors: David Svitov; Pietro Morerio; Lourdes Agapito; Alessio Del Bue
• Keywords: Texture Synthesis

We present billboard Splatting (BBSplat) - a novel approach for 3D scene representation based on textured geometric primitives. BBSplat represents the as set of optimizable planar primitives with learnable RGB textures and alpha-maps to control their shape. can be used in any Gaussian pipeline drop-in replacements Gaussians. Our method's qualitative quantitative improvements over 2D Gaussians are most noticeable when fewer used, achieves 1200 FPS. regularization term encourages have sparser structure, unlocking an efficient compression that leads reduction storage space model. experiments show efficiency standard datasets real indoor outdoor scenes such Tanks&Temples, DTU, Mip-NeRF-360. demonstrate PSNR, SSIM, LPIPS metrics compared state-of-the-art, especially case which, other hand, up 2 times inference speed improvement same rendering quality.