simkoc/Remove360

This repository contains the Remove360 dataset.

The Remove360 dataset provides paired pre/post-removal RGB images and object-level masks captured in real-world indoor and outdoor environments. It is designed to benchmark and evaluate semantic residuals—unintended semantic traces left behind—after object removal in 3D Gaussian Splatting. The dataset covers a broader and more complex range of full-scene representations, enabling comprehensive evaluation of 3D object removal techniques beyond isolated object instances.

This dataset is intended to help assess whether semantic presence can truly be eliminated after object removal and if downstream models can still infer what was removed.

License: CC-BY-NC 4.0

Citing Remove360 Dataset:

If you use Remove360 Dataset in your research, please cite it using the following BibTeX entry.

@article{sam3dteam2025sam3d3dfyimages,
      title={Remove360: Benchmarking Residuals After Object Removal in 3D Gaussian Splatting}, 
      author={Simona Kocour and Assia Benbihi and Torsten Sattler},
      year={2025},
      eprint={2508.11431},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.11431}, 
}