SymSeg: Model-Free Object Segmentation in Point Clouds Using the Symmetry Constraint.

SymSeg is a model-free method for segmenting objects and recovering their symmetries from multiple-view 3D pointclouds. The algorithm takes a pointcloud and an occupancy map of the scene as input and returns a set of object segmentation masks and corresponding symmetries. Please cite the following work if you use it for your research:

Ecins, Aleksandrs; Fermüller, Cornelia; Aloimonos, Yiannis (2018) Seeing Behing the Scene: Using Symmetry To Reason About Objects in Cluterred Environments. International Conference on Intelligent Robots. [Project Page]

This repository contains the C++ implementation of the algorithm and a sample test scene. Additional test scenes can be downloaded from the Cluttered Tabletop Dataset (160 Mb).

Dependencies

• PCL
• OctoMap (with dynamicEDT3D)
• protobuffer

Tested in Ubuntu 14.04 and 16.04 with PCL 1.8.0 and Octomap 1.8.1

Building

To checkout and build symseg in the build directory, execute the following in a terminal:

git clone https://github.com/aecins/symseg.git
cd symseg
mkdir build
cd build
cmake ..
make -j

Segmentation Examples

The examples directory provides examples for three different segmentation modes:

• rotational_segmentation segments rotational objects
• reflectional segmentation segments reflectional objects
• full_segmentation segments rotational objects first, then segment reflectional objects

To segment the provided sample scene execute the following from the bin directory:

./rotational_segmentation ../sample_scene

Symmetry Detection Examples

Additionally we provide code for detecting symmetries in pointclouds of single objects. Note that input pointclouds must contain a single object and must have consistently oriented normals. To detect symmetries execute from the bin directory:

./reflectional_detection ../sample_object_cloud.ply