SteeredMarigold: Steering Diffusion Towards Depth Completion of Largely Incomplete Depth Maps

This is the official implementation of "SteeredMarigold: Steering Diffusion Towards Depth Completion of Largely Incomplete Depth Maps" paper, presented at ICRA 2025.

⚙️ Environment

To run the code, create the following conda environment and set the environment variables.

Create the environment:

conda env create -f environment.yml

activate it:

conda activate steered-marigold

and install pip dependencies:

pip install -r requirements.txt

Set the environment variables pointing to your dataset directory and directory containing results:

export DATASETS_DIR="/path/to/your/dataset/directory"

export MODELS_DIR="/path/to/directory/to/store/results"

🗄️ Data

Download labeled NYUv2 Depth dataset from: cs.nyu.edu. The downloaded .mat file is to be placed in "nyu-v2" sub-directory of your dataset directory.

⚡ Run

To reproduce the results from the paper, run the following scripts:

• ./validate_steeredmarigold_incomplete for steering Marigold with largely incomplete depth maps
• ./validate_steeredmarigold_uniform for steering with depth uniformly sampled from the available ground-truth
• ./validate_marigold Marigold without steering

📚 Citation

@INPROCEEDINGS{steered-marigold,
      author={Gregorek, Jakub and Nalpantidis, Lazaros},
      booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)}, 
      title={SteeredMarigold: Steering Diffusion Towards Depth Completion of Largely Incomplete Depth Maps}, 
      year={2025},
      pages={13304-13311},
      doi={10.1109/ICRA55743.2025.11128449}
}

🖇️ License

Code in this repository is licensed under Apache License, Version 2.0.