Shenglun Chen, Xinzhu Ma, Hong Zhang, Haojie Li, Zhihui Wang. Propagating Sparse Depth via Depth Foundation Model for Out-of-Distribution Depth Completion. IEEE TIP.
2025-11-18 We refactor the code.
2025-8-17 Our work is published.
2025-8-7 Our work is accepted by TIP. The ePrint is available at arXiv.
2025-5-28 We upload the testing dataset.
2025-5-26 We upload the pre-trained models.
2025-5-24 We release the main code. Soon, we will upload the test dataset and pre-trained models.
The main environment requires adherence to the corresponding foundation model.
pip install scikit-image
pip install numba
pip install tensorboardX
pip install omegaconf
pip install opencv-python-headless
pip install matplotlib
pip install h5py
pip install timm
| Model | Foundation Model | Version | Checkpoint |
|---|---|---|---|
| PSD-NK-M | MiDaS | Swin2 Large | Download |
| PSD-NK-D | Depth Anything | Large | Download |
| PSD-NK-Dv2 | Depth Anything v2 | Large | Download |
| PSD-NK-DPr | Depth Pro | - | Download |
| PSD-NK-Mev2 | Metric3D v2 | Large | Download |
| PSD-NK-PDA | PromptDA | Large | Download |
| PSD-NK-UDv2 | UniDepth v2 | Large | Download |
N/K denotes the training dataset NYUv2/KITTI.
Training dataset inclues NYUv2 DC and KITTI DC.
Testing datset includes VOID1500, SUNRGBD, TOFDC, DIML indoor, DIODE indoor, Middlebury, ETH3D, ScanNet, Drivingstereo, Argoverse, Cityscapes, DIODE outdoor, Hypersim, Virual KITTI, KITTI 360, and Stanford2D3D. These dataset can be downloaded from Huggingface. However, ScanNet requires the Terms of Use. The sparse depth maps incorporate four patterns: random mask (1%), random mask (0.1%), pseudo-hole, and Harris corner.
Thanks the authors for their works: CSPN, MiDaS, Depth Anything, Depth Anything v2, Depth Pro, Metric3D v2, PromptDA, UniDepth v2.
@ARTICLE{11125857,
author={Chen, Shenglun and Ma, Xinzhu and Zhang, Hong and Li, Haojie and Wang, Zhihui},
journal={IEEE Transactions on Image Processing},
title={Propagating Sparse Depth via Depth Foundation Model for Out-of-Distribution Depth Completion},
year={2025},
volume={34},
number={},
pages={5285-5299},
doi={10.1109/TIP.2025.3597047}}