Multiview tile-image rendering from 3D Gaussians for 1D camera array

초록

Recently, 3D scene representation based on Gaussian primitives are gaining a lot of research interest due to its usability for real-time photorealistic novel view synthesis via 3D Gaussian rasterization. Since the conventional rasterizer are aimed at rendering single viewport, the rendering pipeline is not suited for rendering multiple viewports simultaneously. In this paper, we propose a method to directly render multiview tile-images from 3D Gaussianbased scene representations. By transforming the pixel coordinates in the CUDA kernel function for pixel value calculation, the proposed rasterizer renders multiple images in a tiled form. In addition, by horizontally shifting the projected images of 3D Gaussians in accordance to the tile index, it is possible to render multiple different viewports. Experiments show that the proposed rasterizer is 4.3-times faster than the sequential rendering of individual tile-images in QHD (2560 1440) resolution. © 2026 IEEE.

키워드