Overview

This repository provides the official implementation of: Leveraging Geometric Priors for Unaligned Scene Change Detection.

We propose a training-free scene change detection framework that leverages geometric priors and foundation models to detect scene changes from unaligned image pairs.

Installation

Step 1: Create Conda Environment

conda create -n geoscd python=3.10
conda activate geoscd

Step 2: Install PyTorch

We use CUDA 11.8:

pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 \
--index-url https://download.pytorch.org/whl/cu118

Step 3: Install Dependencies

pip install -r requirements.txt
pip install git+https://github.com/facebookresearch/segment-anything.git

Download Models and processed datasets

Please download the required foundation models and place them under:

src/pretrained/

Model

Model	Download Link	Target Path
SAM ViT-H	https://github.com/facebookresearch/segment-anything	src/pretrained/sam_vit_h.pth
VGGT-1B	https://huggingface.co/facebook/VGGT-1B/blob/main/model.pt	src/pretrained/model.pt

Datasets

Download from Google Drive: https://drive.google.com/drive/folders/16Z_7EWp--psRxRtgq-SZcoUMOCK8ASuY?usp=drive_link

Example directory structure:

data/
 ├── changesim/
 │── PSCD/
 │── PASLCD/
src/
 ├── pretrained/
 │    ├── sam_vit_h.pth
 │    └── model.pt

Inference

Multi-GPU Inference

cd src

bash scripts/changesim_multi_gpu.sh
bash scripts/paslcd_multi_gpu.sh
bash scripts/pscd_multi_gpu.sh

---

Single GPU Inference

cd src

bash scripts/changesim_single_gpu.sh
bash scripts/paslcd_single_gpu.sh
bash scripts/pscd_single_gpu.sh

---

Evaluation

ChangeSim

python src/evaluations/eval_changesim.py \
--gt-root data/changesim \
--results-root results/changesim \
--output-csv metrics/changesim_summary.csv

---

PSCD

python src/evaluations/eval_pscd.py \
--gt-root data/PSCD/mask \
--results-root results/pscd \
--output-csv metrics/pscd_summary.csv

---

PASLCD

python src/evaluations/eval_paslcd.py \
--gt-root data/PASLCD \
--results-root results/paslcd \
--output-csv metrics/paslcd_summary.csv

---

Acknowledgement

This project builds upon several excellent open-source works:

• GeSCF build robust zero shot SCD upon SAM.

• VGGT provide strong geometric prediction which we use to build robust pixel correspondence and detect occlusion.

• RSCD provide the trianing framework and data augmentation strategy which we apply to other training based method to explore their ability under unaligned settings.

We sincerely thank the authors for their contributions.

---

Citation

If you find this work useful, please cite:

@article{liu2025leveraging,
  title={Leveraging Geometric Priors for Unaligned Scene Change Detection},
  author={Liu, Ziling and Chen, Ziwei and Gao, Mingqi and Yang, Jinyu and Zheng, Feng},
  journal={arXiv preprint arXiv:2509.11292},
  year={2025}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.