DeOcc-1-to-3

Official implementation of the paper:
“DeOcc-1-to-3: 3D De-Occlusion from a Single Image via Self-Supervised Multi-View Diffusion”

---

⚡ TL;DR

DeOcc-1-to-3 is a 3D de-occlusion framework.
From a single occluded image, it synthesizes six structure-consistent novel views, enabling faithful Amodal 3D reconstruction.

🚀 Key Features

✅ Synthesizes six structurally consistent novel views from a single occluded image
✅ Self-supervised training with pseudo-ground-truth views generated by a pretrained multi-view diffusion model
✅ Seamless integration with existing 3D reconstruction pipelines (e.g., InstantMesh)
✅ Introduces the first benchmark for occlusion-aware 3D reconstruction

🛠️ Installation

git clone https://github.com/Quyans/DeOcc123
cd DeOcc123

# create environment
conda env create -f environment.yml
conda activate deocc123

# prepare SAM model
mkdir ckpts
curl -L -o ckpts/sam_vit_h.pth https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

Download pretrained model weight from Google Drive and put it under the ckpts/ directory.

🎆 Gradio Demo

You can quickly try out the model with our interactive Gradio demo:

python app.py

Here is a short video demonstrating how to use the demo:

6.28.mp4

📄 Citation

If you find our work useful, please cite:

@article{qu2025deocc,
    title={DeOcc-1-to-3: 3D De-Occlusion from a Single Image via Self-Supervised Multi-View Diffusion},
    author={Qu, Yansong and Dai, Shaohui and Li, Xinyang and Wang, Yuze and Shen, You and Cao, Liujuan and Ji, Rongrong},
    journal={arXiv preprint arXiv:2506.21544},
    year={2025}
}

🌟 Acknowledgements

Thanks to the following great repositories: Zero123++, InstantMesh