Skip to content

RIC: Rotate-Inpaint-Complete for Generalizable Scene Reconstruction

Notifications You must be signed in to change notification settings

SamsungLabs/RIC

Repository files navigation

RIC: Rotate-Inpaint-Complete for Generalizable Scene Reconstruction

RIC Example

[Project Page] [ArXiv]
Isaac Kasahara, Shubham Agrawal, Selim Engin, Nikhil Chavan-Dafle, Shuran Song, Volkan Isler
Samsung AI Center, New York

RIC Architecture

Table of contents:

  1. Setup environment
  2. Inference
  3. Citation

Setup environment

  1. Make a new directory

    mkdir RIC_Project
    cd RIC_Project
  2. Install depth completion repo from ClearGrasp in this folder

    Follow steps 1, 2, and 4 on their page. (You do not need their data or models)

  3. Clone our repo inside RIC_Project folder

    Navigate back to RIC_Project folder

    git clone https://github.com/SamsungLabs/RIC.git

    Folder structure should look like this:

    RIC_Project                # Root folder
    ├── RIC                    # Code for RIC (This Repo)
    └── ClearGrasp             # Code for depth completion (ClearGrasp)
  4. Install our python dependencies

    cd RIC
    python3.8 -m venv venv
    source venv/bin/activate
    pip install -r requirements.txt
  5. Download weights for depth completion

    Download the weights for surface normal and occlusion boundary estimation and place them in the RIC/weights/ folder.

  6. Install HOPE Dataset

    Navigate to the HOPE dataset page and follow the download instructions.

  7. Add OpenAI API Key

    If you don't have one already, follow these instructions:

    Once you have the key, paste the key on line 28 of "inference_hope_dataset.py"

Inference

  1. To generate predictions:

    python3 inference_hope_dataset.py --data_dir "path/to/hope-dataset/hope_video/" --write_file_root "results/"
  2. To evaluate predictions:

    python3 evaluate_results.py --results_dir "results/"

Acknowledgement

RIC code makes use of ClearGrasp and Dalle-2 implementations. We sincerely thanks the authors for providing their implementations.

Citation

@misc{kasahara2023ric,
      title={RIC: Rotate-Inpaint-Complete for Generalizable Scene Reconstruction}, 
      author={Isaac Kasahara and Shubham Agrawal and Selim Engin and Nikhil Chavan-Dafle and Shuran Song and Volkan Isler},
      year={2023},
      eprint={2307.11932},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

About

RIC: Rotate-Inpaint-Complete for Generalizable Scene Reconstruction

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages