6D Pose Estimation of Objects Using the GOOD descriptor

This repository is the work of Richard Fischer, Bharath Raj, Ruben Spolmink and Bharath Musaudhi Rajan for the course Cognitive Robotics (WMAI003-05.2020-2021.1A) at the University of Groningen.

In this project we try to enhance an existing object recognizer algorithm with an additional feature of orientation validation. Our idea for this implementation is to couple a voting-based point pair feature(PPF) pose estimation algorithm (Drost et al. 2010) with a state of the art object descriptor - Global orthographic object descriptor(GOOD) (Kasaei et al. 2016)

In our approach we use the Restaurant RGB-D Dataset (https://github.com/SeyedHamidreza/restaurant_object_dataset) which has the depth information along with the colour information of pointclouds.

Prerequisites to run the model

• Make sure you have a base point cloud of an object for reference (our default base model is "Knife_Object0.pcd")
• Make sure you have Matlab2019a or later version with the packages (Image processing toolbox, Communication toolbox and Computer Vision Toolbox )

Steps to run the model

• Run "src/rigid3dTransformation.m". Here we read the "Knife_Object0.pcd" and transform it in the X,Y and(or) Z axis rotations. The output image serves as the input to the second script.

• Run "src/PoseEstimation.m". Here we take the base object "Knife_Object0.pcd" and the transformed object as inputs (line 11 and line 72). The model calculates the pose transformation of the second object compared to the base object. The output is displayed in a windowed fashion. A sample output is shown as below

Result of our approach