Ruijie Zhu^1,2, Mulin Yu², Linning Xu³, Lihan Jiang^1,2, Yixuan Li³,
Tianzhu Zhang¹, Jiangmiao Pang², Bo Dai^4
1 USTC ² Shanghai AI Lab ³ CUHK ⁴ HKU
ICCV 2025

                 

The overall architecture of ObjectGS. We first use a 2D segmentation pipeline to assign object ID and lift it to 3D. Then we initialize the anchors and use them to generate object-aware neural Gaussians. To provide semantic guidance, we model the Gaussian semantics and construct classification-based constraints. As a result, our method enables both object-level and scene-level reconstruction.

To train ObjectGS, you should download the following dataset:

• 3DOVS
• LERF-Mask
• Replica
• Scannet++

Or you can use our processed subsets: Google Drive

We organize the datasets as follows:

├── data
│   | 3dovs
│     ├── bed
│     ├── bench
│     ├── ...
│   | lerf_mask
│     ├── figurines
│     ├── ramen
│     ├── teatime
│   | replica
│     ├── office_0
│     ├── office_1
│     ├── ...
│   | scannetpp_ovs
│     ├── 09bced689e
│     ├── 0d2ee665be
│     ├── ...

1. Clone this repo:

git clone git@github.com:RuijieZhu94/ObjectGS.git --recursive

2. Install dependencies:

cd ObjectGS
conda env create --file environment.yml

python ply_preprocessing.py

# 3dgs version
bash train_3d.sh /path/to/your/dataset
# 2dgs version
bash train_2d.sh /path/to/your/dataset

bash render.sh path/to/your/training/folder

python vis_ply.py

Please refer to Gaussian-Grouping for OVS evaluation.

# exporting scene mesh
python export_mesh.py -m path/to/your/training/folder

# exporting object mesh
python export_object_mesh.py -m path/to/your/training/folder --query_label_id -1
## query_label_id: specify a object id (0~255), -1 for all objects

If you find our work useful, please cite:

@inproceedings{zhu2025objectgs,
  title={ObjectGS: Object-aware Scene Reconstruction and Scene Understanding via Gaussian Splatting},
  author={Zhu, Ruijie and Yu, Mulin and Xu, Linning and Jiang, Lihan and Li, Yixuan and Zhang, Tianzhu and Pang, Jiangmiao and Dai, Bo},
  booktitle={International Conference on Computer Vision (ICCV), 2025},
  year={2025}
}

Our code is based on Scaffold-GS, HorizonGS, GSplat and Gaussian-Grouping. We thank the authors for their excellent work!