README.md

March 19, 2026 · View on GitHub

S-VGGT: Structure-Aware Subscene Decomposition for Scalable 3D Foundation Models

Xinze Li¹, Pengxu Chen^1,2, Yiyuan Wang^1,3, Weifeng Su^1,4,Wentao Cheng^1†

¹Beijing Normal University–Hong Kong Baptist University ²Jilin University ³Hong Kong Baptist University ⁴Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science

^†Corresponding Author

Contact: t330026083@mail.bnbu.edu.cn

S-VGGT Overview

📰 News

[Mar 19, 2026] Paper Released on Arxiv
[Mar 18, 2026] Code Released
[Mar 17, 2026] 🎉 S-VGGT has been accepted to ICME 2026.

🔭 Overview

S-VGGT identifies structural redundancy across frames and reorganizes dense scenes into balanced subscenes with a shared reference frame, enabling highly efficient parallel 3D reconstruction with strong acceleration and no loss in fidelity.

S-VGGT Method

⚙️ Environment Setup

First, create a virtual environment using Conda, clone this repository to your local machine, and install the required dependencies.

conda create -n svggt python=3.10
conda activate svggt
git clone https://github.com/Powertony102/S-VGGT.git
cd S-VGGT
pip install -r requirements.txt

Next, prepare the ScanNet dataset: http://www.scan-net.org/ScanNet/

We follow the preprocessing method to ScanNet by choosing 50 scenes following FastVGGT, where we implement a tool to assist you. Please refer to Xinze Li: ScanNet-Process-Python3 for more details.

Finally, configure the dataset path. For example:

parser.add_argument(
    "--data_dir", type=Path,
    default="/home/jovyan/shared/xinzeli/scannetv2/process_scannet"
)
parser.add_argument(
    "--gt_ply_dir",
    type=Path,
    default="/home/jovyan/shared/xinzeli/scannetv2/scannet",
)

Note that, we build our implementation based on FastVGGT, where we observe that the default value of --depth_conf_thresh is too high for VGGT output. We suggest to configure it by a value which is lower that 1.0.

🤖 Usage

ScanNet

Evaluate S-VGGT on the ScanNet dataset with various input sequence lengths.

The --merging parameter specifies the block index at which the merging strategy is applied, which is a parameter from FastVGGT. You can leave it with its default value None to inference and evaluate S-VGGT only, or you can use --merging 0 to use token merging from Block 0 to evaluate the orthogonality of S-VGGT and token-level acceleration method.
The --num_groups parameter overrides the number of subscene groups; None uses auto selection

python eval/eval_scannet.py --input_frame 1000

7 Scenes & NRGBD

Evaluate across two datasets, sampling keyframes every 3 frames:

python eval/eval_7andN.py --kf 3

🍺 Acknowledgements

Thanks to these great repositories: VGGT, Dust3r, Fast3R, CUT3R, MV-DUSt3R+, StreamVGGT, VGGT-Long, FastVGGT and many other inspiring works in the community.
Special thanks to our supervisor Dr. Wentao Cheng for consistent suggestions and efforts to this work.

⚖️ License

See the LICENSE file for details about the license under which this code is made available.