---
license: cc-by-4.0
language:
- en
tags:
- computer-vision
- scene-change-detection
- 3d-gaussian-splatting
- bayesian-inference
- scene-specific
- arxiv:2610.28490
datasets:
- ChamudithaJay/PASLCD
- zr-zhou/CCS3D
---

# Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection

**Changrui Zhu · Simon Julier**  
University College London

[Paper](https://arxiv.org/abs/2610.28490) · [Code](https://github.com/changruizhu96/LEGO-SCD) · [Project page](https://changruizhu96.github.io/work/lego-scd/)

This repository hosts the released **scene-specific fitted parameters for LEGO-SCD**.
The method learns statistical models from automatic reference 3D Gaussian Splatting
edits and unchanged reference observations, then combines current-view evidence
with accumulated history for online scene change detection.

## Contents

| Dataset | Parameter directory | Coverage |
| --- | --- | --- |
| PASLCD | `pretrained/PASLCD/Instance_<1 or 2>/<scene>/` | 20 scene-instances |
| CCS3D | `pretrained/CCS3D/<scene>/` | bedroom, bookcase, desk, livingroom |

Download `pretrained.zip` and extract it to recover the directory layout above.
The uncompressed parameter files total approximately 88 MB. Each scene-instance contains:

- `suppression.pt`: fitted response suppressor.
- `likelihood.json`: image evidence mapping.
- `unchanged.pt`: unchanged-response statistics for historical evidence accumulation.
- `prior.json`: effective pixel prior.
- `contour.pkl`: fitted contour filter for refinement.
- `split.json`: reference-view split information.

These parameters match the reference reconstructions used by the released code.
Unchanged-response statistics are indexed by Gaussian: replacing, pruning, or
reordering a reconstruction requires fitting new parameters. A new scene also
requires its own offline fitting. See the code repository for data preparation
and training instructions.

## Use with the released code

Install LEGO-SCD and prepare the dataset following the
[GitHub README](https://github.com/changruizhu96/LEGO-SCD#install).
The GitHub repository already includes these parameters; this Hub repository
provides an additional download source.

To download this mirror into your LEGO-SCD checkout, run the following from that
checkout in an environment with `huggingface_hub` installed:

```python
from huggingface_hub import hf_hub_download
from zipfile import ZipFile

archive = hf_hub_download(
    repo_id="Teleeeeeeeeeee/LEGO-SCD",
    filename="pretrained.zip",
)
with ZipFile(archive) as parameters:
    parameters.extractall(".")
```

After installation and dataset preparation, use the existing evaluation entry:

```bash
python run_evaluation.py PASLCD
python run_evaluation.py CCS3D
```

Add `--pixel-only` to run without output refinement. The code repository includes
fixed initialization edits and documents online timing separately.

The input datasets, reference reconstructions, and SAM2/XFeat weights are obtained
separately through the code repository's preparation instructions. This Hub
repository contains the fitted parameters, not the full inference environment.

## License and attribution

The project contributors' rights in these fitted parameters are licensed under
**CC BY 4.0**. See [the parameter license](LICENSE.md) and
[the CCS3D notice](CCS3D_LICENSE.md) for attribution and upstream sources.
The source datasets, reference reconstructions, and third-party code and weights
retain their respective licenses.

## Citation

```bibtex
@article{zhu2026learning,
  title={Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection},
  author={Zhu, Changrui and Julier, Simon},
  journal={arXiv preprint arXiv:2610.28490},
  year={2026},
  url={https://arxiv.org/abs/2610.28490}
}
```