ScholarWorks@성균관대학교

초록

Light-field (LF) imaging which captures rich spatial and angular information makes such postprocessing possible as refocusing, depth estimation, and view synthesis. However, substantial storage requirement of raw LF data calls for compact yet high-fidelity LF representations. Although 3D Gaussian Splatting (3DGS) offers a compact representation, its full 3D reconstruction exceeds the needs of LF data in which viewpoints are narrowly constrained. In this paper, we investigate Constrained Viewpoint Gaussian Splatting (CVGS), a 2D adaptation of 3DGS tailored to LF data. We first train the CVGS to represent the central view by a single set of 2D Gaussians and then use it to render the off-central views by predicting how Gaussians are displaced in image space. Experiments on the HCI Light Field dataset demonstrate that CVGS provides a compact yet high-fidelity representation of LF data.

키워드