With advances in extended reality technologies, there is an increasing demand for three-dimensional (3D)
scene reconstruction from captured videos. With the advent of Neural Radiance Fields, research on 3D reconstruction has made rapid progress, leading to environments where 3D reconstruction can be easily achieved only with smartphones. However, editing reconstructed 3D scenes or specific objects remains challenging and is still in its early stages. This study aims to semiautomatically edit 3D scenes reconstructed by Gaussian Splatting. We propose a method to estimate and extract foreground objects automatically from user input, and to cut them out from the scene. By applying to 3D scenes reconstructed from existing trained scenes and captured videos, we confirmed the ability of the proposed method to perform visually plausible cutout of foreground objects.



Name Affiliation Web site
Yuya Matsumoto Keio University




  1. Yuya Matsumoto, Issei Fujishiro: “Semiautomatic cutout of objects from 3D scene reconstructed with Gaussian Splatting,” in Proceedings of Expressive Japan 2024 (The Institute of Image Information and Television Engineers Technical Report, Vol. 48, No. 8, pp.19―22, AIT2024-36), Tokyo University of Technology, March 5, 2024, Excellent presentation awards (in Japanese).
  2. Yuya Matsumoto, Issei Fujishiro: “Semiautomatic cutout of objects from reconstructed 3D scene,” in Proceedings of the 86th National Convention of International Processing Society of Japan, Vol. 2, pp. 201―202 (4P-06), The University of Kanagawa, online & onsite hybrid, March 15―17, 2024, Student encouragement awards (in Japanese).


  1. Grant-in-Aid for Challenging Research ((Exploratory):23K18468 (2023―)

Back to GP team page