Privacy Preserving Structure-from-Motion

This is the COLMAP version for the Privacy Preserving SfM project. The code is heavily based on standard COLMAP, but we removed a lot of code (likely we missed something) that is not directly necessary for the privacy preserving pipeline to reduce the complexity and make the code easier to understand.

Update

We are currently cleaning up the code of the new version that works without known camera calibrations (CVPR 2021). It will be released here soon.

About

In this project we explore privacy preserving methods for SfM by replacing the standard keypoint positions with lines to hide dynamic image content from a processing server. The work was presented at the European Conference on Computer Vision 2020.

If you are using this code in a scientific project please cite

@inproceedings{Geppert2020ECCV,
  author    = {Marcel Geppert and
               Viktor Larsson and
               Pablo Speciale and
               Johannes L. Sch{\"{o}}nberger and
               Marc Pollefeys},
  title     = {Privacy Preserving Structure-from-Motion},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2020},
}

@inproceedings{schoenberger2016sfm,
  author={Sch\"{o}nberger, Johannes Lutz and Frahm, Jan-Michael},
  title={Structure-from-Motion Revisited},
  booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2016},
}

Please note that this is not the exact same code that we used to create the results in our paper.

Running the pipeline

The simplest way to create a privacy preserving model is to add the necessary additional information (namely gravity direction and camera calibration) to the input image folder and then run the pipeline as in standard COLMAP in the GUI (feature extracton, matching, and reconstruction).

Providing the additional information can be done by adding two extra text files per image as in

<image_name>.jpg
<image_name>.jpg.camera_model.txt
<image_name>.jpg.gravity.txt

File formats

Gravity direction

The gravity direction is expected as a space separated text file with a single line, the gravity vector pointing downwards in camera coordinates, e.g.

0.022717 0.999539 -0.020155

Note that we are using the computer vision convention for the coordinate system with the x-axis pointing to the right, y-axis downwards and z-axis to the front (from the viewer into the image plane).

Camera calibration

For the camera calibration we also expect a text file with a single line, containing first the camera model name and then the parameters as comma separated values, e.g.

OPENCV 2575.939766, 2608.293891, 1599.263118, 1257.125449, 0.141865, -0.465301, 0.000000, 0.000000

Please check the camera_models.h file for the supported camera models, their identifiers and the expected parameters.

Installation

As our code is based on COLMAP, the build process is mostly identical.

We deliberately removed the install targets to avoid conflicts with a potentially installed standard COLMAP.

We tested the following steps for Ubuntu 18.04:

1. Make sure you have all dependencies installed

   sudo apt install \
       cuda \
       git \
       cmake \
       build-essential \
       libboost-program-options-dev \
       libboost-filesystem-dev \
       libboost-graph-dev \
       libboost-regex-dev \
       libboost-system-dev \
       libboost-test-dev \
       libeigen3-dev \
       libsuitesparse-dev \
       libfreeimage-dev \
       libgoogle-glog-dev \
       libgflags-dev \
       libglew-dev \
       qtbase5-dev \
       libqt5opengl5-dev 

2. Clone and build ceres

   sudo apt install libatlas-base-dev libsuitesparse-dev
   git clone https://ceres-solver.googlesource.com/ceres-solver
   cd ceres-solver
   git checkout $(git describe --tags) # Checkout the latest release
   mkdir build
   cd build
   cmake .. -DBUILD_TESTING=OFF -DBUILD_EXAMPLES=OFF
   make -j
   sudo make install

3. Clone and build the pipeline

   git clone --recurse-submodule https://github.com/colmap/privacy_preserving_sfm
   cd privacy_preserving_sfm
   mkdir build
   cd build
   cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo
   make -j8

4. Run the gui

   src/exe/ppsfm gui

Datasets

We provide a dataset as an example input on our Project Page.

Problems

We tried to minimize the complexity of the code and test the pipeline extensively, but there is still a high chance that we missed certain cases. If you encounter crashes or other problems while running the pipeline, please let us know by creating an Issue here on GitHub.