PrimiTect: Fast Continuous Hough Voting for Primitive Detection

This repository provides the code accompanying the paper PrimiTect: Fast Continuous Hough Voting for Primitive Detection by C. Sommer, Y. Sun, E. Bylow and D. Cremers, accepted for publication in the International Conference on Robotics and Automation (ICRA) 2020. A preprint can be found on arXiv, and we will provide the DOI once available.

Basic Usage

Dependencies

The code depends on the following third-party libraries:

• Eigen (header-only)
• CLI (header-only)
• Ceres
• tinyply
• CGAL 4.14 (optional, only for comparing against Efficient RANSAC)

Eigen, Ceres and CLI are added to this repository as submodules, and from tinyply, we directly added the source files.

Preparation

• Clone the repository to your computer including all submodules.
• Build Ceres in folder thirdparty/build-ceres-solver/.
• Compile the code using the CMakeLists.txt file:

  mkdir build
  cd build
  cmake ..
  make -j4
  cd ..
  

Parameters

To display all input parameters for the respective executables, call --help in addition to the binary name. In short, these are the most common parameters:

• --folder: the folder containing the input point cloud data
• --in: name of the input point clouds - must by in .ply format for our method, and in .xyz format for CGAL-based code
• --results: the folder in which results shall be written
• --type: primitive object type to be detected, the options are plane, sphere, cylinder, cone, plane-sphere, plane-cyl, plane-cone and all
• --np (default: 4096): number of input points to be used for voting stage, usually 2048-8192 is a good choice here
• --settings (default: all, only needed for our method): choose settings with which our code shall be run, 00 corresponds to the Drost* baseline implementation, while 39 corresponds to our proposed method, including vote spreading, cluster-by-averaging, and weighted averaging for candidate extraction

Run Primitive Detection

Our proposed method

The main file, where our proposed method is run in its most basic form, is cpp/detect_objects/main_voting_sample.cpp. The corresponding executable that is generated is DetectObjects. To run it on a point cloud test_data.ply in folder data, cd into cpp/detect_objects/, and run

bin/DetectObjects --settings 39 --type all --in test_data.ply --folder ../../data --results .

Detection and refinement

The executables of type SphereDetectRefine etc. run a subsequent refinement after primitive detection. Note that this is not part of our proposed method, but we include it to provide some way to get more accurate estimates. It uses C

Refinement of pre-detected primitives

The executables of type OptSphere etc. do only refinement, i.e. they require as input a list of primitive parameters, e.g. obtained from our detection method. Note that these are also only provided for completeness and not part of the method we propose in the paper.

Primitive tracking

In the folder cpp/track_object/, you can find the code we used to track a single geometric primitive in a Redwood sequence, as shown in our supplementary video. The folder contains code for detection with both our method (including the baseline) and the CGAL Efficient RANSAC implementation. If you want the Efficient RANSAC files to build, you will have to uncomment the corresponding lines in cpp/CMakeLists.txt and cpp/track_object/CMakeLists.txt.

Efficient RANSAC for comparison

Finally, we include the code to run Efficient RANSAC on a point cloud in .xyz format, and convert results into our format, for easier comparison. The code can be found in cpp/cgal_test/ and is an adaptation from the CGAL examples on Efficient RANSAC. If you want these files to build automatically, you will have to uncomment the corresponding lines in cpp/CMakeLists.txt.

Data

Redwood data (primitive-rw)

We provide the scripts we were using to obtain the data from the Redwood dataset in the folder data/redwood_lod/. You need to download the .zip folders containing the according sequences first, but then can run extract_first_frames.sh.

The Matlab script RedwoodToPly.m can then convert the depth images to .ply data, as required for our method.

Synthetic data (primitect)

The Matlab script RenderSyntheticAll.m in data/primitect/ will generate random point clouds of the same type as those we were using for our evaluation, and refer to as primitect dataset in the paper.

License and Publication

Our code is released under the GPL (v3+) license, for more details please see the LICENSE file. Also note the different licenses of the submodules in the folder thirdparty.

Please cite our paper when using the code in a scientific project. You can copy-paste the following BibTex entry:

@inproceedings{sommer2020,
    title   = {PrimiTect: Fast Continuous Hough Voting for Primitive Detection},
    author  = {Sommer, Christiane and Sun, Yumin and Bylow, Erik and Cremers, Daniel},
    booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
    year    = {2020}
}