CoRe-GS: Coarse-to-Refined Gaussian Splatting with Semantic Object Focus

Hannah Schieber^1,2,4, Dominik Frischmann^1,2, Simon Boche^1,3, Victor Schaack^1,2, Angela Schoellig², Stefan Leutenegger³, Daniel Roth^1,2

Human-Centered Computing and Extended Reality Lab, TUM University Hospital, Clinic for Orthopedics and Sports Orthopedics, Munich Institute of Robotics and Machine Intelligence (MIRMI)¹

Technical University of Munich²

Mobile Robotics Lab, ETH Zurich³

Friedrich-Alexander University (FAU) Erlangen-Nürnberg⁴

arXiv

Code — Coming Soon

Interested in comparing object removal & extraction?
Contact: hannah dot schieber [at] tum dot de

Abstract

Fast and efficient reconstruction supports time-critical tasks such as tele-guidance and disaster response. CoRe-GS introduces a semantic POI-focused Gaussian Splatting pipeline that first builds a fast segmentation-ready representation and then selectively refines semantically relevant regions. This significantly reduces training time while improving novel view quality in areas of interest.

GAGA

Ours (CoRe-GS)

Gaussian Grouping

Ours (CoRe-GS)

Ours

Architecture

Citation

@misc{schieber2025coregscoarsetorefinedgaussiansplatting,
  title        = {CoRe-GS: Coarse-to-Refined Gaussian Splatting with Semantic Object Focus},
  author       = {Schieber, Hannah and Frischmann, Dominik and Boche, Simon and
                  Schaack, Victor and Schoellig, Angela and Leutenegger, Stefan and Roth, Daniel},
  year         = {2025},
  eprint       = {2509.04859},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2509.04859}
}