TurboSL: Dense, Accurate and Fast 3D by Neural Inverse Structured Light

Project Page

Installation

This project is built on top of the following source code for:

Neural Surface Reconstruction based on Instant-NGP

1. Follow the steps in the repo above for installing PyTorch and tiny-cuda-nn.

2. Install the requirements file with:

pip install -r code/requirements.txt

Dataset

A sample dataset for Structured Light images are provided in data. This includes the raw and undistorted images (in data/images), the projection patterns (in data/patterns), the calibration parameters, and object mask.

Training

To train the model, you can specify the training parameters in code/configs/code/configs/neus-sl-geoalbamb-capture-blur.yaml, and run code/launch.py.

The training outputs will be stored under code/exp/[experiment_id].

The reference training results are provided under code/train_david.

Citation

@inproceedings{mirdehghan2024turbosl,
  title={Turbosl: Dense accurate and fast 3d by neural inverse structured light},
  author={Mirdehghan, Parsa and Wu, Maxx and Chen, Wenzheng and Lindell, David B and Kutulakos, Kiriakos N},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={25067--25076},
  year={2024}
}

Acknowledgments

We thank Yuanchen Guo for their implementation of Neural Surface Reconstruction with Instant-NGP.