GeoPatch Descriptor

This repository contains the original implementation of the descriptor "Learning Geodesic-Aware Local Features from RGB-D Images", published at Computer Vision and Image Understanding journal. GeoPatch is by design invariant to isometric non-rigid deformations of surfaces by leveraging geodesic-invariant sampling, designed as a mapping function before descriptor computation is performed from photometric information.

Learning Geodesic-Aware Local Features from RGB-D Images
[Project Page] [Paper]

Ready to use Docker container

We recommend running the project with Docker, which requires a single command to build the entire project.

First, build the container:

docker compose build

Then, run the docker in interactive mode:

docker compose run --rm geopatch

Finally, you can run the provided demo, which runs both geodesic patch extraction and local feature computation:

sh run_demo.sh

Notice that the output files are being saved inside the container.

Ready to use Singularity container

Alternatively we also provide a singularity recipe so you can easily and smoothly build the project.

First, build the container:

sudo singularity build geopatch.sif Singularity.geopatch

Then, run the container in interactive mode:

singularity shell --writable-tmpfs --pwd /src geopatch.sif

Finally, you can run the provided demo, which runs both geodesic patch extraction and local feature computation:

sh run_demo.sh

Notice that the output files are being saved inside the container.

Datasets

All available datasets and ground-truth files are available for download at https://verlab.dcc.ufmg.br/descriptors

Non-Rigid Simulator

The code for the non-rigid simulator used in our work is available in the nonrigid_sim folder. For detailed instructions on usage, please refer to the README.md inside the folder.

Published works

If you find this code useful for your research, please cite the paper:

@article{potje2022learning,
  title={Learning geodesic-aware local features from RGB-D images},
  author={Potje, Guilherme and Martins, Renato and Cadar, Felipe and Nascimento, Erickson R},
  journal={Computer Vision and Image Understanding},
  volume={219},
  pages={103409},
  year={2022},
  publisher={Elsevier}
}

VeRLab: Laboratory of Computer Vison and Robotics https://www.verlab.dcc.ufmg.br